SCROLL NEWS / DISCOVERY
Search headlines
Find the stories shaping the conversation.
Results for " downstream " 29 found 🔀 AI-powered Shuffle
Politics and Culture Are Downstream of Anthropology and Theology
If Western culture is to survive and flourish, we must recover both the religious and the anthropological vision out of which it arose.
A note to readers: Science, accountability and elections
Dear readers, As early voting begins, I wanted to share some news — and why it connects to the decisions we make at the ballot box. I’ve been selected for the inaugural cohort of the South Atlantic Climate Journalism Hub mentorship program, hosted by Covering Climate Now. I’m excited to join this group, learn from scientists and fellow journalists, and expand my science reporting. Science directly ties to governance. Local officials make decisions about where development goes, how infrastructure is maintained, what emergency preparations receive funding, and how communities rebuild after disasters. Understanding the evidence behind those decisions helps us evaluate whether our government is serving us well. Here in western North Carolina, those connections are all around us, in Lake Lure and beyond. Flood maps inform decisions about property and development. Engineering assessments help identify risks at dams and bridges. Rainfall forecasts guide emergency preparations. Geological information helps explain landslide hazards. Water testing gives communities information they need to assess conditions. Meteorologists’ knowledge and public perception of weather events shape public safety. Each involves scientific or technical knowledge. Each also raises questions about public responsibility: What did officials know? How did they use that information? What did they tell the public? What did taxpayers receive for their money? A study can identify a problem. Elected officials still have to decide what to do about it, how to pay for it, and how to explain their choices. Reporting should examine both the evidence and the response. That is where I want to strengthen my work. My reporting on Helene recovery, infrastructure, flooding, public safety, and public spending has repeatedly brought me to questions that require scientific expertise alongside public records. This mentorship may help me better understand the research, ask more precise questions, and explain the answers clearly. I also bring the perspective of someone living with these concerns. I live downstream of Lake Lure Dam, and landslides affected my property during Helene. Those experiences bring urgency to my questions. Careful reporting helps establish the answers. Readers deserve to know what the evidence shows, where uncertainty remains, and whether officials accurately describe the information guiding their decisions. You also deserve scrutiny of proposed solutions: what they cost, what they are expected to accomplish, and what risks remain. Access to that information matters. As a mostly deaf journalist who relies on written communication, I will continue pressing for written public warnings, records, and explanations that people can read and use in real time. As you consider the candidates seeking your vote, consider how they approach evidence, preparedness, and transparency. Do they support local journalism, even when it’s not parroting a press release? Do they explain weather events clearly in plain English and explain their decisions after them? Make supporting records available? Acknowledge uncertainty? Follow through when a report identifies a risk? Those are practical questions about governing. Thank you for supporting Cops & Congress and the work that makes these questions part of the public conversation. I’m grateful to David Dickson and the CC Now team for this opportunity, and I look forward to bringing what I learn back to you. Annie DanceEditor & Publisher, Cops & Congress — Learn more about this newsletter, ethics policy, how you can help shape this work, and support it. Follow on X and Facebook. Send constructive criticism, fan mail, and tips with public documents for future stories: CopsandCongress@gmail.com ICYMI: View the Cops & Congress archive — Annie Dance is the publisher of Cops & Congress, a newsletter that analyzes what happens when crime, courts, disaster, democracy, and small-town policies collide. Views expressed here are covered by the First Amendment. Dance has a Bachelor of Arts from Manhattan University in Communication with a focus in Journalism and Government. She has been a journalist for over 20 years. Share Upgrade now or give a gift subscription
USTR Seeks Business Input on the EU Carbon Border Adjustment Mechanism
The USTR is seeking comments from stakeholders on the current and potential effects of the EU’s CBAM on U.S. producers, workers, exports and overall U.S.-EU trade. USTR is specifically requesting information on costs, regulatory burdens, EU market access, competitive effects and the possible impact of proposals to extend CBAM to additional downstream products
Community Bridges built ECM to prevent health crises. Now the state funding is shifting and we have to close the program
California created enhanced care management to help Medi-Cal patients with complex health and social needs before they reach the emergency room. Community Bridges spent three years building a community-based program that now serves more than 200 people. But, Community Bridges CEO Ray Cancino writes, a shift toward encounter-based reimbursement is making the relationship-building, outreach and coordination that prevention requires increasingly difficult to sustain and Community Bridges will now be forced to close the program. Cancino wonders here the new payment model will actually save money — or simply move the costs downstream.
Faraday Future Announces This Week’s Investor Weekly Report From YT Jia Has Been Rescheduled to September 28; FF EAI “Built in USA” Upstream & Downstream Business Partner Conference Part Two to Be Held September 29
FF Founder and Global CEO YT Jia’s Investor Weekly Report has been rescheduled for the afternoon of September 28 and will be released at its usual time. FF has invited 10 global partners to share insights on the EAI robotics industry, FF’s “Built in USA” strategy and related topics, highlighting the potential for collaboration between […]
Why Is BP (LSE:BP.) Reshaping Its Business And Investment Approach Now?
BP (LSE:BP.) CEO outlined a new investment approach focused on more selective capital allocation across projects. Management is reorganising BP into separate upstream and downstream divisions to increase operational accountability and clarity. The group adjusted fuel deliveries and production in response to the recent Iran conflict, highlighting its operational flexibility. This new upstream and downstream structure, combined with tighter project selection, only captures part of what is...
Science Vs on the 49ers Substation: Anatomy of a Hatchet Job
In August, Pablo Torre Finds Out and Science Vs released an episode about my theory that electromagnetic fields from the electrical substation beside the 49ers facility may be contributing to the team’s run of soft-tissue injuries. As the originator of the theory, I was also a subject of the episode. I sat was interviewed by Wendy Zukerman and Rose Rimler of Sciene Vs. I sent them research, connected them with scientists, and spent hours digging through the City of Santa Clara’s and Silicon Valley Power’s records to help clarify critical details. Both Wendy and Rose came off as affable and professional, showing curiosity during the interviews, with Wendy going out of her way to question the “conspiracy theory” framing itself. Then they titled the episode We Solve the 49ers Viral Conspiracy Theory. I was familiar enough with their catalog that I never expected them to agree with my theory, but I did expect them to give it a fair hearing and to deliver their findings honestly. On both counts they came up short. They were not directly dishonest in anything they said, but rather in what they left out, in how what remained was cut, and in the picture painted by those two things in combination. They asked me to recorded the interview on my phone for better audio quality, and I held on to my copy for reference later. That turned out to be a fortuitous decision. Here is the raw audio of my first interview, unedited*. Only my side was recorded, so their questions aren’t on it, but you can generally infer them from the answers, and every claim I make below about what I said can be checked against my statements in this clip. (Note: Wendy had connection trouble at the beginning of the interview and was cutting in and out, which accounts for the false starts.) 0:00 -1:19:13 Audio playback is not supported on your browser. Please upgrade. Science Vs is a fun show. It is pitched at roughly a high-school understanding of science, for an audience that is skeptical, Snopes-oriented, and worried about misinformation. Take a claim circulating online, interview a supporter of the claim, interview consensus scientists, and then tell people which side of the line it lands on. I have listened to their episodes on raw milk, seed oils, 5G and other contested issues in the wellness space. I have problems with all of them, but each at least made an honest attempt to represent the actual range of scientific opinion by talking to credentialed scientists with differing perspectives. Dissenting researchers were heard from, answered, and often given a chance to respond. This episode did not do that. This article is not a relitigation of the science. I have written extensively about that already. This post is for the record. It is simply an accounting of what was wrong with their episode. Contents Is there a power plant next to the 49ers practice facility? — the clearest single example of getting something objectively wrong For the record — a high-level overview of all of the problems with their investigation Draw your own conclusions — why they had to make this episode at all What they left out about childhood leukemia — the exposure threshold, and playing games with statistics to minimize sick children What they left out about the mechanism — the body of literature they refused to cover, and evidence for the mechanism used against it The scientists Science Vs interviewed and did not air — four credential scientists who support myhe theory, recorded and cut, or not even called How they misrepresented me and my case — an answer severed from its question, and three telling edits How you would actually study the injuries — it doesn’t actually matter which team is the most injured, overall The theory: magnetic fields, oxidative stress, and collagen Since the episode never states my theory, here it is below. Those who are familiar with my work can skip this section. Electrical equipment produces magnetic fields, and those fields induce faint electrical currents in anything that passes through them, including people. My argument goes further than induced current. Mitochondria — the structures that generate a cell’s energy — run on continuous flows of electrons and protons, and those flows carry faint electromagnetic fields of their own. Maxwell’s principle of superposition shows that when two fields share the same space, they add together. An external field does not need to be strong to alter an internal field, it only needs to be present. The electron transport chain inside the mitochondria leaks a small fraction of its electrons by design, and those become reactive oxygen species — free radicals, which healthy cells use as signals. External magnetic fields cause the rate of leakage to change, increasing free radicals. When free radicals accumulate faster than the cell’s antioxidant systems can clear them, the result is oxidative stress. Oxidative stress is a recognized driver of chronic illness, including cardiovascular disease, neurodegeneration, diabetes and cancer, and it is in every medical school curriculum. One of its ordinary effects in connective tissue is that it switches on matrix metalloproteinases — MMPs, the enzymes that break collagen apart — while suppressing new collagen synthesis. Tendons are made of collagen. An athlete loading his soft tissue heavily six days a week is already creating oxidative stress in the tendons, so the addition of magnetic field exposure compounds the damage and inhibits repair, making him more susceptible to injury when enormous force lands on a tendon in a fraction of a second. That is the theory. We know that EMF causes oxidative stress, and we know that oxidative stress degrades collagen. The only thing missing is a study showing that direct connection. I have never claimed certainty that the substation is contributing to the elevated injury rate. I do believe it is likely, and I believe it warrants caution, and that a revenue machine like a major NFL team can easily afford to remove the exposure out of precaution. Is there a power plant next to the 49ers practice facility? Here is the clearest single instance of the problem with this episode. It requires no knowledge of electromagnetic fields at all. This story first came to me through Jon Feliciano, a former 49er, in an interview on Willard & Dibs on 95.7 The Game. Asked about why the 49ers are always injured, he said players used to joke that it was the power plant behind the practice fields. Wendy Zukerman corrects him on air. “It’s important to say that it’s not a power plant near the 49ers site.” Torre then asks her to explain the difference between a substation and a power plant. Her answer: a power plant generates electricity and transmits it to a substation, which converts the voltage into something your house can use and sends it on to your home or your business. There are two problems with that. The first is that there is a power plant on that site. The Gianera Generating Station is a natural-gas-fired peaker plant owned by Silicon Valley Power, and it has been there since 1986. I have published aerial photographs of the location from 1987 in which the gas plant is the only thing standing, before the practice facility was even there. I even mentioned the gas plant in my first interview. (Feliciano has since gone on record about this issue, and it is clear he had the substation in mind rather than the gas plant.) The second is that her definition describes a distribution substation, and the facility beside the 49ers is not that. It is the Northern Receiving Station, which takes in bulk power from outside the city’s grid and delivers it to distribution substations. Silicon Valley Power’s own map distinguishes receiving stations from substations, and Rose Rimler even reads the label out loud in the episode when she walks through that map on air. A few minutes before the map segment, Zukerman tells Torre: “like I said, we fact check everything.” Well, maybe not everything. I made a mistake On the morning of July 6th, at 10:20, I sent Rose and Wendy two studies. During the interview I had been asked which human studies show oxidative stress from electromagnetic fields, and I had not been able to name one on the spot. Rose emailed asking for the studies, I quickly searched for a couple, and I wrote that I thought these were the most relevant from that search. Unfortunately, I was busy and did not read either one closely at first. I picked them because they were topically relevant — one on substation workers, one on power plant workers. Neither were great studies, and their problems go past the ones raised in the episode. That was careless on my part. There is context that did not make the episode, though. Before that email I had been on a Zoom call with Rose talking about EMF and oxidative stress, and I pointed her to Dr. Henry Lai at the University of Washington, who has spent years aggregating the literature on this question. His tallies show that roughly 90% of studies since the late 1990s find oxidative effects from electromagnetic field exposure — hundreds of studies, systematically counted. Rose told me she was already aware of his work. So when she asked for studies, I understood her to have the aggregate view already and that she would report anything within that context. Not reading those two carefully was still a mistake, but the larger mistake was assuming the episode would be based on a thorough review of the actual body of literature, rather than on whatever I happened to send that morning. The second email At 1:49 that same afternoon — three hours and twenty-nine minutes after the first — I sent them a second email that I had been preparing for a while. It was a structured document in six numbered sections, written in response to a request they had made two weeks earlier, laying out my entire argument with citations attached to every link in the chain. The episode aired on August 13th, five weeks later, and not a single item from the second email is in it. Here is how it opens: The full email is included at the bottom of this post in Appendix B. Nearly 1,500 words in total. It contains two named mechanisms with the physics behind them. Occupational studies in humans. The literature on oxidative stress and collagen. A section addressing the strongest counter-evidence against my own theory, which I raised myself. And a closing section stating plainly what the theory cannot yet show. My entire argument, in writing, with detailed citations for every claim, provided to them five weeks before broadcast, and none of it made it into the episode. That is the single largest omission, but it is not an isolated one. The whole episode is riddled with similar problems, and the conclusion they came to is not merely a difference of opinion about the scientific evidence. Through omission and through strategic editing, Science Vs misinformed its audience — about what the evidence shows, about what I argued, and about who they spoke to. That is a notable outcome for a show whose stated purpose is correcting misinformation. For the record I am not going to tell you that Science Vs set out to discredit me, or that they had come to their conclusion before they reported the story, or that they used misleading editing to make me look like a man retreating from his own theory. I cannot know what was in anyone’s head, and I am not going to pretend otherwise, either. What I can tell you is what is on the record. They asked me for evidence and aired the weakest thing I sent. They interviewed three scientists who supported my theory and aired none of them. They discussed childhood leukemia without ever saying the exposure level at which it appears. And they took an answer I gave weeks before Ken Karipidis, the one EMF scientist who was put on air, was interviewed, and placed it directly after his critique, where it comes off as a major concession. A friend of mine who has worked as a journalist for thirty years read that last one and called it sleazy. I agree with her. Those are just the four issues that are impossible to defend as pure editorial decision, but they are not the only issues. The most glaring issues stem from what they left out: The exposure level at which the childhood leukemia association actually appears — three to four milligauss, a number the episode never states despite it being much lower than what I measured. The mechanism itself: The episode gets as far as “oxidative stress could weaken collagen,” which leaves it sounding like speculation, rather than a known pathway. The second email, and all the studies in it. Dr. Paul Héroux at McGill, whose work my proposed mechanism rests on, whose contact details I supplied, and who was never called. Dr. Beatrice Golomb at UC San Diego, who runs a research group on oxidative stress and mitochondrial function and has specific expdertice in collagen, interviewed and not aired. Dr. David Carpenter at Albany, the most prominent academic voice on the harm from magnetic field exposure, interviewed and not aired. Dr. Joel Moskowitz at Berkeley, interviewed and not even credited. The specific, testable condition I put in writing that would make me abandon the theory. My own unprompted statement of the theory’s evidentiary gap. My caveats about my own measurements. My ranking of EMF as one factor among several rather than the only one. The new equipment installed at those facilities between 2010 and 2014, and the sharp rise in system load from 2013 onward. But that is not all. What they kept, they selectively edited to serve a narrative that was not served by the raw transcripts of the interviews. A description of thousands of commercial routers in a workplace, reduced to the Wi-Fi in your house. My own acknowledgment, multiple times, that "I'm not a scientist" clipped in an episode that spends considerable energy pointing out that I am not a scientist. And a qualification I volunteered about my own measurements, cut from my answer and delivered later by their EMF expert as a correction. Together their edit produce a portrait of a wellness guy with Wi-Fi phobia and two bad studies, backing away from his theory the moment a “real” scientist looked at it. A practice facility where nothing changed because no new substation was built. Injuries written off as random because the 49ers were not quite the single most injured team in the league over the period in question. That portrait is false even though almost no individual statement made in the episode is. Draw your own conclusions At the top of the episode, Pablo Torre says his show had been flooded with requests to investigate the substation — that it was the thing they were asked to look into over and over again. Neither show went hunting for this story. It arrived from their own audience, in volume, and it could not be waved off. A thread I posted on X reached twenty-two million people. A guy with a gaussmeter and a brand-new Substack obligated a Pultizer prize winning program to spend an hour on a question about player health, that is considered fringe enough that most media operations won’t touch it. Pablo Torre Finds Out could not ignore it. So Science Vs came in and ran a “debunking” operation instead of an investigation. Everything below is what had to be left on the cutting room floor to deliver the “debunking”. They could not ignore it. So they tried debunked it instead. The demand was strong enough to force an episode, and since the episodes release the 49ers have sustained more injuries, including season ending ones, and interest has only grown. Demand from fans and playets is strong enough to force a better investigation from someone else. An omission is not a lie until you find out what was omitted. What follows is the detailed record so you can see the full picture. Draw your own conclusions. I have drawn mine. I am going to keep publishing. A note about Pablo Torre Finds Out After the episode aired I posted publicly, calling out Pablo Torre Finds Out for concerns I had about the integrity of their investigation, and about an interview directly with Pablo Torre which had been offered but never happened. I have since spoken with a producer at PTFO. The offer had rested on a misunderstanding about where things stood between us, and it fell away as the Science Vs part of the production widened past the science. He apologized for the miscommunication and I accepted. The producer also confirmed to me that they did essentially hand off the entire investigation to Science Vs for narrative purposes — so Pablo would be genuinely curious or potentially surprised by their findings — while still taking responsibility for what went out under their name. I appreciated the candor and the accountability, and I am comfortable with where we left things. So this post is not about PTFO. My disagreement is with Science Vs. What they left out about childhood leukemia Childhood leukemia appears at three milligauss. They gave no number at all. 1. They built an EMF ruler, then discussed childhood cancer without it. The sequence on air went like this: First I measured almost 9 mG, then I measured up to 50 milligauss. Torre reacts with concern — “that sounds high.” Cut to Karipidis: the public limit is 2,000 milligauss. Torre reacts again — “What happens at 2,000?” Karipidis explains that nothing happens at 2,000, that safety factors are built in. Then: at more than 50,000 milligauss you get flickers of light in the retina, and higher still, nerve stimulation. The listener now has a ruler (their ruler) to measure these numbers with. Fifty is nothing by their ruler. Then the episode turns to childhood leukemia and states, accurately, that there is a real association — but tellingly, they never give the exposure level at which that association appears. Childhood leukemia appears at three milligauss. They conveniently left that out. The pooled analyses that produced the World Health Organization’s carcinogen classification found roughly a doubling of childhood leukemia risk above three to four milligauss. That is the number that belongs in that segment, and it is more than ten times below the reading that had just been dismissed as trivial. I gave them that number myself, on tape, in the interview. Describing my own threshold, I said my comfort level for long-term chronic exposure is three to four milligauss, but preferably one. 2. Six children in thirteen million. The episode reports that these fields might explain one to four percent of childhood leukemias, then reframes it as a German estimate of perhaps six cases among thirteen million children. Torre, to his credit, is uncomfortable. What never gets explained is why that number is small. Almost no children lives above three milligauss in Europe. Fewer than two children in a hundred, in the Italian data. So even if the risk to those children doubles, the number of additional cases across an entire country stays small — not because the hazard is small, but because hardly anyone is exposed to it. Count the whole population and the few who are exposed disappear. And the reason hardly anyone is exposed is that siting rules keep housing away from transmission lines. In Europe, at least. The Dutch government has a standing precautionary policy against building new homes in any zone where the annual average magnetic field exceeds 4 milligauss — the same number the episode never mentioned. Denmark and Norway treat 4 milligauss as the level that triggers a look at mitigation. That is five hundred times below the ICNIRP figure quoted on air. The United States has no federal equivalent. That is evidence of infrastructure siting policy working. It is not evidence that there is no hazard. It is evidence that we already act as though there is one. If you listen to the raw audio you will hear that my proposed solution is EMF exposure is to build better infrastructure. What they left out about the mechanism 3. They accused me of cherry-picking, and then cherry-picked themselves. On air, Zukerman tells Torre that people who worry about electromagnetic fields dismiss the studies showing null results and collect the ones showing harm, and that this is why you have to be careful not to cherry-pick. They had two papers I had rattled off in a hurry, and a second email containing a carefully chosen set. They ran the two weakest, graded them on air, and presented the result as the state of the evidence for not only my theory, but the whole question of oxidative stress from magnetic fields. That is cherry-picking from my emails. The bigger picture is what they did with the literature as a whole. Zukerman acknowledges it exists, dismisses it, and disposes of it in a single stroke. You can go online right now, she tells Torre, and find scores of studies showing these fields increase oxidative stress. The issue, she says, is that they are not studies in people. Many of them are in people. The occupational studies in my second email are in people. That email had been in her inbox for five weeks. She waved off the literature on a factual claim that the documents in her possession contradict. (I have written separately about the evidentiary standard at work here). The effect was that they essentially baited me: ask the subject to supply the evidence, take the weakest thing he sends, and let that stand as a proxy for the whole argument. Wendy has been doing this show for ten years and clearly knows how to extract what she needs for her desired narrative. 4. They presented the collagen-improvement research as a gotcha, but I had already addressed it. The episode treats the existence of studies showing that EMF can improve collagen as a surprise that undercuts me, but I had pre-emptively addressed that five weeks earlier in the email they ignored. Those studies use measured, high-intensity pulsed dosing. The mechanism by which they work is oxidative signalling. The cell reads a controlled rise in reactive oxygen species as a warning, upregulates its antioxidant defenses, mounts a repair response, and the therapy stops before the system is overwhelmed. The key point is that it induces an oxidative response, which means those studies are actually evidence for my proposed mechanism, not against it. A field that produces a therapeutic effect cannot be biologically inert (which is the implication when they say “Nothing happens at 2,000 milligauss). Chronic ambient exposure is the same pathway with no off switch, and it is most consequential during the evening and while sleeping, when the body is in repair mode and the exposure does not stop. Section 5 of my email is titled What about the study showing improved collagen synthesis. I raised the counter-evidence myself, explained why I did not think it refuted the theory, and wrote that I was disappointed the question had not come up in the interview. That is worth holding against the cherry-picking charge in the item above: the studies they presented as inconvenient for me were studies I had gone out of my way to put in front of them. 5. Section 1 of my email stated a critical argument that was never put to the expert directly responsible for it. My email says the exposure limits exist to prevent acute effects and do not address chronic exposure at all, and that my dispute is with the entire framework. Zukerman gestures at this once. After Karipidis gives the 2,000 milligauss figure, she tells Torre that I would object because those levels do not capture “more subtle effects.” That is not my argument, and “subtle” is doing a lot of work here. My argument is that the limits are built to prevent acute effects — nerves firing, muscles twitching, etc — and were not designed to address chronic exposure at all. In fact, ICNIRP’s own documents state as much themselves. Mine is a structural objection to what the standard measures, reduced as a concern about subtlety, it becomes a matter of degree, rather than a categorical difference. They interviewed Ken Karipidis, vice chair of ICNIRP — the body that writes those limits. He was asked what the limit is, but he was never asked how it applies to chronic exposure. The scientists Science Vs interviewed and did not air 6. They interviewed four officials of the standards body. The episode credits roughly a dozen researchers. Four of them are associated with ICNIRP. Ken Karipidis is the vice chair. Rodney Croft chaired ICNIRP from 2020 to 2024 and led the development of its current guidelines. Dan Baaken is ICNIRP’s scientific secretary. Frank de Vocht sits on the Main Commission. One of them speaks in the episode. Their institution’s position is more than adequately represented — the 2,000 milligauss figure, the safety factors, the assurance that nothing happens below the line. 7. The four who were left out. Dr. Beatrice Golomb is Professor of Medicine at UC San Diego — a physics undergraduate degree, a PhD in biology, an MD. She directs a research group whose stated focus is the relationship between oxidative stress, mitochondrial function and health. That is the mechanism the episode spent its first half dismissing. Her published work includes the fluoroquinolone case series behind the FDA’s black box warning on an antibiotic class that ruptures tendons, where the accepted mechanism is oxidative stress and mitochondrial toxicity — the back half of my chain, already on a drug label. She is in the credits. She is not in the episode. She has told me she finds my theory biologically plausible, and that she has a recording of her interview. Dr. David Carpenter directed the Institute for Health and the Environment at the University at Albany for decades and is the most prominent academic voice arguing these fields cause harm. Childhood leukemia is the reason these fields carry a WHO classification at all, and Carpenter is the person you call if you want that case made by someone who has spent a career on it. He is in the credits. He is not in the episode. He has told me that when he was interviewed they could not have been less interested in what he had to say. Dr. Joel Moskowitz directs the Center for Family and Community Health at UC Berkeley’s School of Public Health. He told me directly that they interviewed him and that he thought it went well. He does not appear in the episode, and unlike Golomb and Carpenter, he does not even appear in the credits. Dr. Paul Héroux at McGill is the researcher whose work provides the foundation for the oxidative stress mechanism my theory rests on. I gave Rose his email address and phone number. He was never interviewed. They never even called him. 8. “He’s not a scientist.” The episode leans hard on the fact that I am not a credentialed scientist. Zukerman says it. Torre repeats it. “At all.” Having built a segment on that, they then left out the physicist-biologist-physician, the professor of public health, the director of a university environmental health institute, and the McGill researcher whose work the primary mechanism comes from. They spent the episode telling their audience I am not a scientist, and then kept four actual scientists who support the theory off because their argument does not survive otherwise. How they misrepresented me and my case 9. They played my words as a response to an interview I had never heard. Near the end of the hour, Zukerman says she went to me and told me the evidence was lacking. Then she plays me saying I am not wedded to it being an EMF, that I have staked my reputation on it needing more study, and that I am not certain of anything. That segment was placed immediately after Karipidis had finished taking the cherry-picked papers apart. It plays as a man folding under expert scrutiny. Commenters have used it against me since, but it is incorrect. Here is the sequence as it happened. I was interviewed. Karipidis was interviewed weeks later. I was never told what he said, or even who he was. I was never asked to respond to it. I did not know his name until the episode aired. That is not a simple editing decision. That is a quote from one interview, moved to follow a critique from a different interview it was not a response to, so that it would land as a concession. A friend of mine who has spent thirty years as a journalist read that sequence and called it sleazy. She is right. The full passage is reproduced in Appendix A, and it is worth reading. What airs is the preamble. Four paragraphs later I say “to directly answer your question” — and that answer, about what would change my mind if the team relocated, is not in the episode. Neither is the sentence two paragraphs on, where I say the EMF might turn out to be ten percent of the problem. I will say the underlying point again, though, because I meant it and I still do. I am not wedded to my theory that magnetic fields from the substation are contributing to the 49ers injuries. I could be wrong. What I am wedded to is the process of scientific inquiry, and the fact that we should be cautious around these exposures and we need more independently funded studies. 10. Three telling edits. They cut “I’m not a scientist” mid-sentence. What I said: “...and I am not certain of anything. I’m not a scientist. To me, I find the mechanisms compelling. I will go back to the fact that I got into this not because of the 49ers, but because of clients I was working with who suffer from EDS...” What aired: “...and that I am not certain of anything. I’m not.” The clip ends on the first two words of the sentence in which I say I am not a scientist, in an episode that spends considerable energy establishing that I am not a scientist. The full sequence is in Appendix A. They cut my own description of my work. Asked what I do, I said I run a wellness business, that I am focused on writing software to help people navigate environmental stressors, that environmental health is my focus, and that I also do independent journalism and research. What aired is the wellness business and the environmental health, with the software and the journalism removed from my own answer. The software engineering does come up later — narrated by them, as a hacker who grew up to write firmware — but my current work, in my own words, arrives as “wellness guy.” Thousands of commercial routers became home Wi-Fi. What I described: working at a wireless networking company surrounded by thousands of Wi-Fi routers, and a stress-induced health collapse afterward that working in that environment contributed to. What aired: Wi-Fi, the kind you have at home — reinforced by a clip of Christian McCaffrey talking about turning his router off at night. Three cuts, and every one of them makes me seem less credible. That is not a coincidence. 11. I gave them my own caveats and the test that would disprove me. In the first email I wrote that a well-designed chronic, ambient-level exposure study in tendon showing no collagen degradation, run over many months, would undercut my confidence in the theory. That is a specific, testable condition, stated in advance and in writing. There was a second interview, weeks later, which was not used in the episode, but in that I asked again for the falsification criterion to be included. What the episode presents instead is a man saying “we need more study,” which reads as a hedge, an open-ended request for the benefit of the doubt. It is the opposite of a hedge. Section 6 of my second email says, unprompted, that as far as I am aware there is not a single study showing EMF-driven oxidative stress degrading collagen in a human in vivo or in human tendon in a lab. I wrote that because it is true and because it is the first thing anyone should check. Under their standard it becomes the hinge of the entire first half, framed as something they had to surface without my cooperation. I had handed it to them five weeks earlier. And in the interview I said that my 25 to 50 milligauss readings were taken at knee level, and that these fields attenuate quickly, so by the time it reaches your head it is not that high anymore. That did not air either. The qualification arrives later, from Karipidis, as a correction. How you would actually study the injuries 12. They measured the wrong thing. The episode spends its final act on whether the 49ers are the most injured team in the NFL, finds no clean ranking, and lands on “injuries are random”. That is the wrong test. You would define the exposure — measured field levels at the facility — and the outcome narrowly: soft-tissue injuries, not a single MSK injury count that a broken finger and a ruptured Achilles both land in. Then you would compare the 49ers to themselves before and after the exposure changed, and compare that change against every other team over the same window. That controls for league-wide shifts in rules, turf and training load. It is a difference-in-differences design, and it is standard. I sent them that argument on July 28th, in writing, along with the fact that in eleven of the past twelve seasons the 49ers had finished top ten in Adjusted Games Lost. And by the episode’s own reporting, the 49ers are second in games lost to injury over the past decade, with FTN’s Aaron Schatz confirming that once you adjust for starters they come out second worst in the league — after having been one of the healthiest teams for years before that. A finding like that is where an investigation starts, not where it stops. 13. I did a lot of substation research for the episode that was never aired. I had reported that the substation next to the facility was expanded in 2014. That was wrong, and Rose Rimler and I sorted it out together. They gave me credit on air, which I appreciated, but framed it as a concession on my part to something they had found. In fact I had been investigating it myself, was uncertain before they asked, but was waiting for clarification from Silicon Valley Power. The uncertainty centered around a substation named “Esparanca” that was proposed right behind the facilities as part of the stadium construction, but now appears to never have been built. I spent days in Silicon Valley Power’s records, and did eventually reach Silicon Valley Power for confirmation and indentified the mistake myself. I also gave them highly detailed documents on what we do know about those facilities: new equipment installed between 2010 and 2014, and load growth across the system with an inflection around 2013–2014, with the Northern Receiving Station carrying the highest capacity in the system and sitting closest to the data center corridor. None of that is in the episode. I have a separate piece about it coming out soon. *I cut 2 seconds from the raw audio to remove personal information that had no connection to anything in the episode. Appendix A: the passage the clip was taken from This is the sequence from my first interview, in full. The episode uses one fragment of it. It’s tricky because my understanding is that there’s something like a 15-year half-life on collagen remodeling issues. They’re going to get a new crop of players. What I really need, though, is for the injuries to be broken down so we can see the soft tissue injuries. Right now, they’re categorized as MSK injuries, but let’s say we can actually identify them more specifically. Is that right? He was able to do that? Oh, cool. I wasn’t able to find that information. Interesting — I didn’t know that data existed. I tried so hard to get it, and I didn’t know it was available. Maybe you could connect me with that person. That would be amazing. I’m not wedded to it being an EMF, honestly. I’m not. I would say that, to a certain extent, I’ve staked my reputation on the fact that it needs more study. It’s important to look into it, and I am not certain of anything. I’m not a scientist. To me, I find the mechanisms compelling. I will go back to the fact that I got into this not because of the 49ers, but because of clients I was working with who suffer from EDS, Ehlers-Danlos Syndrome, a tendon hypermobility issue. This condition is documented to be partially caused by oxidative stress. I’ve worked with them to mitigate their EMF exposure and have seen people improve. I think it’s a bigger story. In a way, the 49ers represent an occupational cluster with specific exposure risks. If someone would conduct a study on this, I believe we could learn a lot. Perhaps we would discover that it’s like, okay, maybe it’s 10% of the problem. I don’t know the exact number, but I think either way, conducting the study would provide valuable insights. To directly answer your question, it’s challenging for me because it’s part of a bigger story. If the 49ers move and stop having injuries, I might conclude that it was related to the substation. However, if they move and still experience injuries, I would want to analyze the nature of those injuries. Are they primarily soft tissue injuries? I don’t think it’s necessarily a case closed, but the story might lose traction at that point regardless. Three things are worth noting. The aired fragment stops immediately before the sentence in which I say I am not a scientist — the cut described in item 10. It also stops four paragraphs before the words “to directly answer your question.” What the episode plays is the throat-clearing. The actual answer, about what would change my mind, never airs. And two paragraphs in, I say the EMF might turn out to be ten percent of the problem. That is the ranking the episode says I never gave. Appendix B: the full email, July 6th, 1:39 PM Hi Rose, You asked me to send you some studies a couple of weeks ago. I apologize for the delay. I have been incredibly busy. I also wanted to give you a chance to dig in yourself, and get your own perspective before I influenced you too much with mine. That said, I do want to make sure that the actual argument is abundantly clear, so here it is! One note before I begin, there are really two separate questions here: Does the substation cause the soft-tissue injuries? That was my specific claim, and what the podcast is about. Is the theory biologically sound? Can 50/60 Hz fields at these levels do anything biological below the safety limits? Question two is why people have been calling me a conspiracy theorist. The regulatory and advisory bodies (ICNIRP, etc) maintain that anything below their exposure limits is biologically inert, and that position is broadly accepted within the mainstream, but the actual science paints a different picture. To defend my theory I’ve had to defend a much larger claim: that the advisory bodies have the biology wrong. I can’t honestly answer question 1 without first addressing that. Here’s is a simplified outline of my theory, with links to studies: 1. Advisory body thresholds for biological effects The ELF magnetic-field limits (ICNIRP’s ~2,000 mG public / 10,000 mG occupational) exist to prevent acute effects from exposure. If a magnetic field is strong enough, it makes nerves fire and muscles contract. Muscles twitch. You can see faint flashes of light in your vision. People can even feel tingling or even a shock. That happens due to Faraday’s law, which states that a time-varying magnetic field induces an electric field inside the body via the creation of Eddy currents.The standards assume that those currents do nothing below the twitch threshold. They do not consider chronic effects at all. My argument is that they are biologically active and do have chronic effects. 2. There are multipe ELF-EMF mechanisms for inducing oxidative stress. This is basic biophysics. Below are the two primary mechanisms. Simplest first. Dr. Paul Héroux of McGill has the most intuitive explanation, which rests on Maxwell’s principle of superposition: because Maxwell’s equations are linear, two sets of electromagnetic waves sharing the same space simply add together. In the interview I pictured a pebble dropped into a pond, sending out clean ripples. Then a boat motors past at distance, sending larger waves, and where they meet the ripples, the two add together and the original pattern gets distorted. In our cells, the mitochondria run on continuous currents of electrons and protons that carry their own faint electromagnetic fields. An external field from a substation, by the principle of superposition, adds to those internal currents and shifts the pattern. The electron transport chain in the mitochondria leaks a fraction of its electrons by design as reactive oxygen species (ROS), which act as signalling molecules in a healthy cell. Distorting that field increases that electron leakage, creating oxidative stress. A tiny electromagnetic interference results in a large downstream effect. https://www.sciencedirect.com/science/article/pii/S2405844025006474 The more complex pathway is Dimitris Panagopoulos’s Ion Forced-Oscillation / Voltage-Gated Ion Channel (IFO-VGIC) mechanism. Weak, polarized, low-frequency fields force the ions (calcium, potassium, sodium) next to an ion channel’s voltage sensor to oscillate; because that force scales with the inverse cube of the distance involved, fields far below the stimulation threshold can still force the channel open, disrupting the cell’s ion balance and driving ROS overproduction (Nitric oxide and superoxide interact and produce peroxynitrite, a highly unstable cytotoxin that breaks down into free radicals). https://www.researchgate.net/publication/366088934 https://www.spandidos-publications.com/ijo/59/5/92 https://www.mdpi.com/1422-0067/22/18/10041 https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2025.1585441/full https://pmc.ncbi.nlm.nih.gov/articles/PMC3772193/ 3. Evidence of oxidative stress, in vivo, in humans from ELF-EMF There are many studies, and the strongest are occupational, which applies to the 49ers situation. 115 power-plant workers vs. 145 office controls. Serum malondialdehyde (MDA, the standard lipid-peroxidation marker), superoxide dismutase, and catalase were all significantly higher in the exposed group, and the markers rose with increasing field exposure. https://journals.sagepub.com/doi/abs/10.3233/WOR-203244 A follow-up where they reversed the oxidative stress in power plant workers. A double-blind randomized controlled trial: 91 power-plant workers given vitamin E, vitamin C, both, or placebo for 90 days. The antioxidant groups showed reduced lipid peroxidation and raised antioxidant capacity versus controls. https://pubmed.ncbi.nlm.nih.gov/32191586/ A spot-welder study: 46 welders occupationally exposed at 50 Hz, showed significantly altered red-blood-cell antioxidant enzyme activity and concluded ELF-MF “might act as an oxidative stressor… even at the recommended levels of exposure.” https://pubmed.ncbi.nlm.nih.gov/18504600/ A substation-specific study found oxidative stress and DNA damage https://pubmed.ncbi.nlm.nih.gov/24460415/ 4. Oxidative stress degrades collagen This half isn’t controversial at all. Oxidative stress activates the matrix metalloproteinases (MMPs) that break collagen down and suppresses its synthesis. Here are three example studies: Siwik et al. 2001: Oxidative stress raises MMPs and lowers collagen synthesis in cardiac fibroblasts. https://pubmed.ncbi.nlm.nih.gov/11121376/ Fisher et al. 2009: Collagen fragmentation drives oxidative stress and elevates MMP-1 in aged human skin. https://pmc.ncbi.nlm.nih.gov/articles/PMC2631323/ Varani et al. 2006: Decreased collagen production in oxidatively aged skin. https://pubmed.ncbi.nlm.nih.gov/16723701/ 5. What about the study showing improved collagen synthesis I’m glad you asked this question, and I’m bummed we didn’t get to it in the interview because it is a great question, and very important to understand. Superficially it may appear to be a refutation of my theory, but I would argue that it actually supports the underlying argument. That study used a short, high-intensity pulsed dose: therapeutic-PEMF, the same type of field used clinically to stimulate bone and tissue healing. The reason this works is because it triggers the very same biological pathways that can lead to oxidative stress. The exposure still stimulates ROS production, but done at a controlled dosage, for a limited duration it has a different effect. The cell reads the ROS as a warning signal and it upregulates its antioxidant defenses, and mounts a repair/proliferation response. The therapy is then stopped before the cell gets overwhelmed. That is essentially a hormetic response: a small dose of a stressor producing a beneficial adaptation. Two things follow: At the very least, this means the field is biologically active, which by itself undercuts the regulatory claim that these fields are inert below their thresholds. There is a wealth of literature on PEMF (pulsed EMF) bone regeneration and calcium channel signalling (which is the second mechanims in section one). Here is one. https://www.nature.com/articles/srep13856 At best, it leads us straight to my actual concern, which is that under chronic, chaotic and unregulated exposure, the cell’s antioxidant capacity gets overwhelmed, resulting in oxidative stress. There are plentiful examples of things that are beneficial at controlled doses, but harmful at uncontrolled doses. Excecise is one obvious, uncontroversial example. Botox is perhaps a more interesting example. 6. What about a study that shows EMF causing collagen degradation in humans? As far as I am aware there is not a single study showing EMF-driven oxidative stress degrading collagen in a human, in vivo, or in human tendon in a lab. There are studies involving isolated fibroblasts though, which secrete collagen. The most relevant one showed that human fetal scleral fibroblasts exposed to 50 Hz ELF-EMF showed a significant decrease in collagen synthesis and that MMP-2 expression was upregulated, which is the exact mechanism I described in section 2. https://pmc.ncbi.nlm.nih.gov/articles/PMC3626379/ That’s it! Thanks again for the interview. Best, Peter Anthony Cowan Environmental Health · Wellness Tech · Researcher & Journalist Publisher & Editor, Living Energy Bio: petercowan.comSocial: X · Instagram · TikTokApp: sunlightis.lifePhone: [redacted]
Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading
Engram extends standard token embeddings with learned multi-token lookups. Recurring local patterns retrieve vectors directly, reducing the need to reconstruct them through attention and feed-forward layers. With Engram model architecture optimization, it allows for lower HBM capacity to be needed for models at the same quality. This does not mean there won’t be an insane demand for HBM but it just means that model architecture will continue to innovate around constraints. This model architecture design is naturally codesigned for parameter offloading: each token accesses a few embedding rows whose addresses depend on token IDs, not hidden states. The runtime can prefetch those rows from host DRAM while earlier layers compute, keeping the table outside HBM without transferring entire weight matrices. Our Memory model contains our latest estimates of quarter by quarter HBM, DRAM, & NAND supply and demand. Offloading frees HBM for model weights and KV cache, potentially supporting larger batches or more concurrent sessions. When DRAM becomes the next constraint, NVMe offers another tier. Recommendation systems already cache frequently or recently accessed embedding rows in faster memory while backing colder rows with SSDs. After NVIDIA roadmap had to change due to massively despec’ing Rubin Ultra from 1024GB to now ~200GB of HBM per chip, model architecture optimizations like emgram maybe helpful. Our DeepSeek-V4.1-Flash configuration uses roughly 189 GiB of memory for Engram. We replace it with a memory-mapped (mmap) file and measure serving performance with offload to SSD. Later on in our report, we will show our Engram offloading experiments along with the official InferenceX agentic inference serving results on Engram models like DeepSeekv4.1 Flash across all 6 NVIDIA GPU SKUs along with MI355X. Unsurprisingly, the CUDA Moat is still mogging MI355X on the ultra popular DeepSeekV4.1 Flash model. We also show how even on high capacity HBM SKUs, offloading emgrams to DRAM could result in even better performance for most of the pareto than keeping the emgram in HBM. Our benchmark has been widely reproduced, validated and/or supported by almost every major buyer of compute from Google Cloud to Microsoft Azure to Oracle, to Meta and many more. Furthermore, it has the support of the ML community including from vLLM, LMCache, SGLang, PyTorch, Huggingface and the support of major labs like OpenAI, MiniMax, ZAI, Qwen, Moonshot Kimi, etc. Star the InferenceX GitHub repository if you find the open-source benchmark and data useful!. InferenceX is the only inference benchmark in the world to have TPUv7, Jalapeño, Nvidia Rubin NVL72, AMD, and soon, SambaNova and Trainium. Due to how realistic AgentX scenario is to real world agentic inference workloads, AMD has committed to collaborating on MI455X UALoE72 too. DeepSeek did not release the original paper’s two trained Engram models. We replicated its setup on fineweb-edu using the released code and training hyperparameters, at an estimated 6E18 FLOPs per run. We observed the same U-shape scaling: Engram improved performance over pure MoE baselines. We also reproduced DeepSeek’s results, where earlier-layer representations with Engram resembled those of later layers. Like the original Engram paper, it is possible to probe Engram’s gate scores to see what n-grams DeepSeek-V4.1-Flash makes the most use of. Our gate scan finds names, code fragments, relational phrasing, and boilerplate. These examples prioritize interesting-ness over gate strength. One unexpected result was Wright : Ace Attorney. These examples suggest learned memory optimizes the training objective, not a judgment of which facts deserve storage. Licenses, bibliography fragments, API scaffolding, and website furniture can provide prediction shortcuts, so the value of additional Engram capacity may depend on what survives data preparation. This does not show that table capacity is “wasted”: the evaluation-corpus scan establishes neither training exposure nor the capacity occupied by each category. For offloading, strong gates do not identify cache-hot rows. Low gates do not automatically save reads either: computing the gate requires the retrieved key and undoes the performance gain of a fused kernel. Skipping reads would require a separate usefulness predictor before retrieval. In the original paper’s inference-time ablation, factual-knowledge benchmarks retained just 29–44% of their original performance, while reading comprehension retained 81–93%. This is due to the training–inference mismatch. The resulting degradation therefore measures this trained model’s dependence on Engram, not the performance difference between models trained with and without it. In our ablations, suppressing Engram worsens token likelihood across all evaluated domains, especially encyclopedia text and several code corpora. Surprisingly, GSM8K accuracy stays within measured run-to-run variation and removing Engram has no effect. Engram is not a detachable dictionary beside an unchanged MoE. Removing it changes downstream features and expert selection. We tested whether rerouting hurts or compensates by holding tokens fixed in a teacher-forced experiment on CRUXEval, a code-reasoning benchmark of small Python functions where the model predicts a function’s output from its code and an input, and scoring the reference answer. Removing Engram raised answer loss from 0.2848 to 0.3093 bits/token. Forcing the ablated model to use the original Engram-on expert choices made it worse still, at 0.3375 bits/token. Rerouting partially compensates for the missing memory. Memory features and expert selection work together, rather than following a clean “memory stores facts; experts reason” division. On the same CRUXeval, removing Engram during either phase reduced accuracy and increased generated tokens; removing it throughout produced the largest changes. Keeping Engram for prefill leads to more correct answers than keeping for decode likely due to semantically richer KV cache transferred to decode workers, allowing it to mitigate some of the performance loss. Engram’s table is large, but each lookup is small. DeepSeek-V4.1-Flash requests 24 rows at each of two Engram layers, about 12.4 KiB per processed token position across the model, or 3.1 KiB per GPU when split across four GPUs. Currently as of Day 7 since Model Release, MI355X is still 2-4x worse performance per dollar compared to B200 even when normalized by Mi355X’s lower TCO. Our full total cost of ownership breakdown comes from our AI Cloud TCO Model along with monthly market surveys of over 100+ gpu clouds & gpu cloud customers. On the Day 0 release of DeepSeekv4.1 Flash, NVIDIA vLLM works out of the box with zero issues across all 6 SKUs: H100, H200, B200, B300, GB200, GB300! This was thanks to the amazing work by the NVIDIA & Interact teams! In comparison, AMD vLLM did not work on day 0 for DeepSeekv4.1 Flash.Source: SemiAnalysis InferenceX AMD’s vLLM documentation points to using vllm/vllm-openai-rocm:deepseekv41-flash-0909, but from hour 0 of the model release to hour 23, AMD has not publicly released the image. AMD claims, “SPEED IS THE MOAT,” yet it has still not released it by the 23rd hour. We wish that, going forward, the AMD team has a better process for hour 0 model releases. Eventually when they did publicly release for “day 0” image support, performance-wise, it is currently up to 14.8x worse perf per dollar than H200 and up to 42x worse perf per dollar than B200/B300. The power of the CUDA MOAT is NVIDIA’s collaboration with its massive 6 million-developer community ecosystem including most of the vLLM & SGLang & Tokenspeed maintainers which means that CUDA is optimized on day 0. Overall, AMD did make significant improvements but the performance per dollar is still currently 2-4x worse than B200. AgentX Engram DRAM Offloading Improving Performance The HBM and DRAM offload use the same GPU kernel to select and dequantize rows. With HBM, it reads GPU memory; with Unified Virtual Addressing (UVA), it reads pinned host memory directly. Both support full decode graphs. Moving the table into HBM accelerates only the sparse lookup, leaving decoder computation and communication unchanged, resulting in little overall benefit while consuming memory otherwise available to KV cache. Another benefit of Engram offloaded to DRAM which means you can reduce the communication overhead by using less HBM GPUs per replica. For example, when enabling Engram offloading on B300, we are able to switch from TP4 to now TP2 which improves the pareto curve by up to 1.6x. When iso-model quality, less HBM is required as the engrams could be offloaded to host DRAM. Thus HBM bandwidth matters way more than HBM capacity. For inference workloads where memory bandwidth matters the most, 4-hi HBM provides the best $/bandwidth and therefore lowest cost per token. If China continues to make more and more revolutionary model architecture innovations, soon it could potentially 0Hi HBM stacks. [ ](https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi) [ Long Live the Short King: Why 4-hi HBM Wins ](https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi) Myron Xie, Bryan Shan, and 3 others · Sep 13 [ Read full story ](https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi) Moreover, on B300 and week-0 stack, moving the Engram table back to HBM did not improve results and stay within run-to-run variance. This is a result of the work optimizing DRAM offload, such as async, and overlap. SSD Offloading On B200, we created a unoptimized vLLM fork and stored the Engram tables in memory-mapped files on local SSDs. File backing lets the OS reclaim table pages when other applications need RAM. Pages already cached in memory can be served without reading the SSD again. Furthermore, note that we were unable to turn on GDS The unoptimized SSD implementation changes how rows reach the GPU. It copies row IDs to the CPU, deduplicates them, gathers the requested rows into pinned buffers, copies those rows back to the GPU and dequantizes them. This work runs between segments of the GPU execution graph. Native UVA performs row selection and dequantization directly on the GPU, avoiding the CPU round trip. A warm filesystem cache removes physical SSD reads, but leaves the coordination, row gathering and transfers. This is why a file already cached in RAM can still perform worse than a pinned DRAM table. The comparison measures the whole serving path; it does not separate the time spent on each of these operations. B200 DRAM dominates both measured SSD serving curves in total tokens per dollar and P90 interactivity. Near 125 tokens/s/user, DRAM delivers 121 million total tokens per dollar versus 52 million for SSD. For production serving, SSD offloading is likely not worth the tradeoff. On the B200 configurations we measured, SSD offloading loses on both measures: every observed SSD point has a DRAM alternative that delivers higher P90 interactivity and more total tokens per dollar. The only points where Cheaper storage does not automatically produce a cheaper inference service. Moving Engram to SSD leaves the same four expensive GPUs and the rest of the server in place. Reclaiming RAM only creates an economic benefit if it enables a cheaper server configuration or additional useful capacity. The current unoptimized path provides neither benefit, and the filesystem cache still consumes RAM when table pages are resident. Next we will look at the mechanisms and specific implementation of ngrams from DeepSeekv4.1 Flash, LongCat, Qwen3.8 Flash Next.
Did New Film Unit Just Shift AMC Stock’s Investment Narrative?
AMC Entertainment Holdings recently launched Leawood Films, a film distribution unit based in Los Angeles and Leawood, Kansas, aiming to bring more small and mid-sized movies into theatres worldwide while relying on fully financed projects to limit production risk. The new distributor uses AMC’s excess screen capacity and past concert-film marketing experience to secure theatrical placement for filmmakers who might otherwise struggle to find backing, while leaving most downstream revenue and...
Colorado Department of Health releases testing on Arkansas River
Leaders with the Department of Agriculture said they have alerted farmers in communities downstream of the spill and are working to get more information from testing.
Public health alert: Sewage spill contaminates Arkansas River downstream of Pueblo
Health officials are warning the public to avoid the Arkansas River downstream of Pueblo – until further notice – after a “significant accidental release” of untreated sewage from the city of Pueblo’s wastewater treatment plant last week was found to have contaminated the water with high levels of E.coli. “The contaminated water will move downstream, […]
Reimagining how scientists study the chemistry of the air
Professor Victoria Barber and doctoral student Cooper Crabtree are spearheading Keck Foundation-funded project charting a path toward better understanding atmospheric oxidation. by Álvaro Castillo, UCLA College of Letters and Science From the vantage point of Victoria Barber, a UCLA assistant professor of chemistry and biochemistry, understanding the chemical reactions that shape air quality and climate are among the most pressing scientific challenges of our time. And thanks to a recent grant from the W.M. Keck Foundation, she and doctoral student Cooper Crabtree will spearhead a new research project exploring exactly how pollutants form in the atmosphere. Specifically, the project will examine how volatile organic compounds transform through atmospheric oxidation and explore how pollutants, like smog, form in the atmosphere. VOCs are released by sources ranging from trees and wildfires to vehicle exhaust pipes and consumer products such as cologne and perfume. As they undergo atmospheric oxidation, they can transform into a wide range of other compounds. Anyone in Southern California can attest to the urgency of this area of research — for example, in late July of this year, Los Angeles County was under an active ozone pollution advisory for almost a full week. While many people associate ozone with the protective ozone layer high in the stratosphere, ground-level or tropospheric ozone is a harmful secondary pollutant formed through atmospheric oxidation. Understanding how VOCs transform into pollutants like ozone is therefore critical to improving air-quality forecasts and informing strategies to reduce exposure. “My group is focused on trying accelerate our understanding of these networks of reactions so that we can do a better job of predicting how those pollutants form,” Barber said. “When we make decisions about how we might alter emissions, we can understand how that’s going to influence the production of these downstream pollutants.” With support from the W. M. Keck Foundation, Barber and Crabtree are developing new experimental approaches that will combine — and improve — techniques to study hundreds of chemical systems more efficiently than traditional methods allow. Instead of spending years decoding a single VOC, the project seeks to dramatically expand the pace and scale of discovery. The grant comes at a particularly important moment for environmental research. Barber, who joined UCLA in 2023, says funding for climate and air-quality research, especially for early career scientists, has become increasingly difficult to secure, a challenge compounded for ambitious projects that require building entirely new research infrastructure. The Keck Foundation’s support, Barber said, provides the flexibility to pursue bold early-stage, high-risk ideas that may be more difficult to fund using conventional federal funding sources. “It’s really, really hard to do environmental research under the current federal paradigm,” Barber said. “By stepping in and providing stable and flexible funding to explore this new idea, the Keck Foundation is making a huge difference to my research program.” In addition to funding early-career faculty, the Keck Foundation’s support is also focused on empowering graduate students like Crabtree in their scientific growth. The project gives Crabtree the opportunity to help develop an entirely new experimental approach to studying atmospheric chemistry — an experience that will provide him with valuable technical skills, scientific training and preparation for the next stage of his career. Stable funding also allows doctoral students like Crabtree to devote themselves more fully to research rather than balancing their scientific work with additional teaching responsibilities needed to cover funding gaps. “Getting to devote more of my time to research this past school year has greatly improved the pace and quality of our experiments, data analysis and planning for our future work,” said Crabtree, who is pursuing a Ph.D. in chemistry. “I feel very lucky — especially amid the current uncertainty surrounding scientific funding — to be able to continue exploring exciting knowledge gaps in our field and growing as a scientific researcher.” For Barber, mentorship is among the most rewarding and important aspects of her role as a faculty member. “Our jobs as faculty are to do good science and make good scientists,” she said. “I believe the Keck Foundation shares this belief and sees the potential positive impact this project can have.”
Business in Oak Creek Canyon worried about impact of flooding threat
Last week’s flooding turned Oak Creek dark as water carried charred debris downstream after the Pocket Fire. Forest Houses Resort says floodwaters rose about 5 feet at its crossing, leaving it impassable until crews cleared sediment. Owners say continued flash flood concerns have prompted guest cancellations, even as their private property remains accessible. Businesses say open trails, restaurants and other attractions still give visitors reasons to come to Sedona and Oak Creek Canyon.
Abu Dhabi oil giant snaps up Shell’s South African business in $1billion Africa expansion
ADNOC Distribution will acquire Shell’s South African downstream business for $1 billion, giving the Abu Dhabi energy group control of 580 service stations and expanding its African footprint
Five Million Americans Lost ACA Health Insurance—Here’s What That Could Mean For Public Health
New data shows 5 million Americans have lost health insurance through the ACA. A doctor breaks down the downstream consequences for public health in America.
Regenerative Agriculture and Population Health
Regenerative agriculture can enhance food quality and reduce harmful exposures through improved soil biology, biodiversity, and ecosystem health, yielding downstream benefits for agriculture, food systems, and human health. It may improve profitability through greater land resilience and lower input costs and may deliver benefits to farms, farmers, communities, and the broader environment. More research is warranted to understand the full range of potential health benefits of regenerative agriculture.
BP's new upstream unit to be led by Gordon Birrell, downstream by Richard Harding
BP said on Tuesday that company veteran Gordon Birrell would lead its upstream unit and Richard Harding will be interim head of the downstream unit as part of its previously announced organisational change into two main business segments.
Raizen to sell Argentina downstream business to Mercuria in $1.42 billion deal
Brazilian sugar and ethanol producer Raizen said on Thursday that its subsidiary Raizen Energia S.A. had signed a binding agreement to sell its downstream operations in Argentina to Mercuria Energy Group Ltd. for $1.42 billion.
Legal victory for home and business owners impacted by Hurricane Harvey flooding downstream from Addicks and Barker dams
A federal judge ruled downstream property owners near the Addicks and Barker dams suffered damage after government water releases following Hurricane Harvey.
'Coyote vs. Acme' & Other Ketchup Movies To Be Sold By V10
Ketchup has entered into a distribution partnership with V10 for the latter to sell the pay-1 window and downstream rights to Ketchup's movies.
V10 Entertainment To Handle Pay 1 & Downstream Sales On Ketchup Movies Including ‘Coyote vs. Acme’
EXCLUSIVE: Ketchup Entertainment has entered into a distribution partnership with V10 Entertainment for the latter to sell the pay-1 window and downstream rights to Ketchup’s movies, including the upc...
V10 Entertainment To Handle Pay 1 & Downstream Sales On Ketchup Movies Including ‘Coyote vs. Acme’
EXCLUSIVE: Ketchup Entertainment has entered into a distribution partnership with V10 Entertainment for the latter to sell the pay-1 window and downstream rights to Ketchup’s movies, including the upc...
Advancing Mental Health Worldwide: Five Core Elements for Effective Strategies
More than one billion people worldwide live with a mental health condition (MHC). Despite the significant downstream economic, societal, and health challenges a
Affordability and Preventive Public Health Policy
Affordability pressures increasingly shape health risk in the United States, influencing both the upstream conditions that sustain health and the downstream ability to access health promoting resources. Financial stability is a key driver of health, affecting patterns of health, health care use, and the tradeoffs people must make among competing needs. The economic policy landscape aimed at improving financial security for Americans is expansive, complex, and often difficult to organize, making it challenging to discuss how different policies influence financial resilience and population health. We propose the Earn–Keep–Grow framework as a practical way to organize and guide discussion of these policies in population health research and policy decision-making.
Gemma 4 and what makes an open model succeed
Having written a lot of model release blog posts, there’s something much harder about reviewing open models when they drop relative to closed models, especially in 2026. In recent years, there were so few open models, so when Llama 3 was released most people were still doing research on Llama 2 and super happy to get an update. When Qwen 3 was released, the Llama 4 fiasco had just gone down, and a whole research community was emerging to study RL on Qwen 2.5 — it was a no brainer to upgrade. Today, when an open model releases, it’s competing with Qwen 3.5, Kimi K2.5, GLM 5, MiniMax M2.5, GPT-OSS, Arcee Large, Nemotron 3, Olmo 3, and others. The space is populated, but still feels full of hidden opportunity. The potential of open models feels like a dark matter, a potential we know is huge, but few clear recipes and examples for how to unlock it are out there. Agentic AI, OpenClaw, and everything brewing in that space is going to spur mass experimentation in open models to complement the likes of Claude and Codex, not replace them. Especially with open models, the benchmarks at release are an extremely incomplete story. In some ways this is exciting, as new open models have a much higher variance and ability to surprise, but it also points at some structural reasons that make building businesses and great AI experiences around open models harder than the closed alternatives. When a new Claude Opus or GPT drops, spending a few hours with them in my agentic workflows is genuinely a good vibe test. For open models, putting them through this test is a category error. Something else to be said about open models in the era of agents is that they get out of the debate of integration, harnesses, and tools and let us see close to the ground on what exactly is the ability of just a model. Of course, we can’t test some things like search abilities without some tool, but being able to measure exactly the pace of progress of the model alone is a welcome simplification to a systematically opaque AI space. Share The list of factors I’d use to assess a new open-weight model I’m considering investing in includes: Model performance (and size) — how this model performs on benchmarks I care about and how it compares to other models of a similar size. Country of origin — some businesses care deeply about provenance, and if a model was built in China or not. Model license — if a model needs legal approval for use, uptake will be slower at mid-sized and large companies. Tooling at release — many models release with half-broken, or at least substantially slower, implementations in popular software like vLLM, Transformers, SGLANG, etc due to pushing the envelope of architectures or tools. Model fine-tunability — how easy or hard it is to modify the given model to your use-case when you actually try and use it. The core problem is that some of these are immediately available at release, e.g. general performance, license, origin, etc. but others such as tooling take day(s) to week(s) to stabilize, and others are open research questions — with no group systematically monitoring fine-tunability. In the early era of open models, the days of Llama 2 or 3 and Qwen pre v3.5, the architectures were fairly simple and the models tended to work out of the box. Some of this was due to the extremely hard work of the Llama, Qwen, Mistral, etc. developer teams. Some is due to the new models being genuinely harder to work with. When it comes to something like Qwen 3.5 or Nemotron 3, with hybrid models (either gated delta net or mamba layers), the tooling is very rough at release. Things you would expect to “just work” often don’t. I’ve been following this area closely since we released Olmo Hybrid with a similar architecture, and Qwen 3.5 is just starting to work well in the various open-source tools that need to all play nice together for RL research. That’s 1.5 months after the release date! This is just to start really investing more into understanding the behavior of the models. Of course, others started working on these models sooner by investing more engineering resources or relying on partially closed software. The fully open and distributed ecosystem takes a long time to get going on some new models. All of this is lead-in for the most important question for open models — how easy is it to adapt to specific use-cases? This is a different problem for different model sizes. Large MoE open-weight models may be used by entities like Cursor who need complex capabilities in their domain, e.g. Composer 2 trained on Kimi K2.5. Other applications can be built on much smaller models, such as Chroma’s Context-1 model for agentic search, built on GPT-OSS 20B. The question of “which models are fine-tunable” is largely background knowledge known by engineers across the industry. There should be a thriving research area here to support the open ecosystem model. The first step is to understand characteristics of different base and post-trained models to understand what they look like. The second step is to tune pretraining recipes for open models so they’re more flexible. Interconnects AI is a reader-supported publication. Consider becoming a subscriber. Subscribe For The ATOM Project and other Interconnects endeavors, we’ve put in substantial effort to measuring adoption trends in the open ecosystem. Everything takes a long time to unfold after a model is first publicly available — and adaptability is why. What we know for sure now, when Qwen has been going from strength to strength with its releases, is that technical staff across the industry has gotten comfortable working with Qwen models. Countless research methods and datasets were made to work with Qwen. It’ll take patience for any other model family to get to this point — a patience I’m not sure many open model builders have. This takes us to Gemma 4, Google’s latest open models. Gemma 3 was released more than a year ago, in March of 2025, and is a bit underrated. Gemma 4 comes in 4 sizes for now, with a bigger, MoE model of over 100B total parameters rumored but not released yet. The models we have today come in sizes of ~5B dense, 8B dense, 26B total 4B active MoE, and 31B dense. I’m most excited that they’re finally adopting a standard Apache 2.0 open source license. This’ll massively boost adoption. The standard of better licenses for strong open-weight LLMs was set by mostly Chinese open model labs in the last 1-2 years, and now U.S. companies are following suit. I will personally be so happy if the horrible Llama licenses and Gemma terms of service were an ~18-month transient dynamic of the industry being nervous about releasing strong open models. The Gemma 4 scores look very solid, the small models have incredible benchmark scores (especially in general domains like LMArena) and the 31B model rivals the recent Qwen 3.5 27B, which is the leading member of that class. The ~30B size range is an important one, as it’s accessible both to researchers and to enterprises looking to deploy the model in real use-cases. Where the 7B model scale is the default for tinkering and research, a 30B model is the default for seeing if an open model can unlock substantial value in your specific workflow — a good mix of intelligence, low price, tractability for downstream training, etc. This takes us back to the above adoption criteria I mentioned for open models and the bigger question — do I think Gemma 4 will be an overwhelming success? Previous Gemma models have been plagued by tooling issues and poorer performance when being finetuned. Gemma 4’s success is going to be entirely determined by ease of use, to a point where a 5-10% swing on benchmarks wouldn’t matter at all. It’s strong enough, small enough, with the right license, and from the U.S., so many companies are going to slot it in. I’m cautiously optimistic that Gemma 4 is going to work better here. Winds are shifting for open models built in America. We saw GPT-OSS go through a bumpy launch to become an overwhelming success. There’s a collective energy around the likes of Reflection, Arcee, Nemotron, Gemma, Olmo, and peers that show substantial demand for building new stacks around open models. There’s capital to be spent on AI stacks across the economy by those who want more ownership of everything, including the model. After launching The ATOM Project 240 days ago, the conversation is shifting into the next stage. Summer of 2025 was a crisis moment where the U.S. AI scene realized it can’t wait and figure out open models after building AGI. The two markets will capture different areas and proceed in parallel. Now that more companies in the U.S. are releasing strong models, we need to improve the ecosystem so that these models are easy to use, understand, and build value around. It’s the hard work to build another inflection point in these adoption plots I’ve been updating consistently, but that’s the work to be done. Join me in it. More data coming soon! Here’s a sneak peek:
Split Decision: Matthew Liberatore is Evolving
Does a new splitter from Matthew Liberatore have unexpected downstream effects that will carry over to the season?
'Culture is downstream from entertainment' — An Interview with Rob Schneider
'You have a left that is totalitarian in the sense that they’re not for free speech. They don’t really have the best interests of the culture in mind. They’re not open for debate. They’re open for demonization. They want to demean you, they want to demoralize you, they want to prevent you from speaking, and they want to take away your ability to make a living.'
GYESME: A New Design-Led Downstream Of GNOME Being Explored
A new project trying to get off the ground and currently in an 'exploratory phase' is GYESME that describes itself as a 'design-led' downstream of GNOME with plans ot only fork when needed that is 'minimal by default.'