SCROLL NEWS / DISCOVERY
Search headlines
Find the stories shaping the conversation.
Results for " OpenAI " 818 found 🔀 AI-powered Shuffle
September Monthly Review: Health MCP Guy?
For all the time I spend talking about APIs and data exchange, this newsletter has had embarrassingly little interoperability of its own. It’s old-school SaaS, baby - UI or bust. Not all of that is my fault! Substack, for all its popularity, is (as mentioned before) really frustrating as my system of record. The primary user it optimizes towards is the reader, whose attention feeds the app, the Notes feed, and the recommendation network that Substack is actually selling. Health API Guy is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Subscribe They’re falling into the same trap that Medium once sprinted into, albeit at a much more glacial speed. Younger generations may not remember (or never knew) it, but Medium once lured publications like The Ringer and Signal v. Noise onto its platform, only to lose them as it rebuilt itself around a paywall and an algorithmic feed. Substack’s original pitch circa 2020 was the antidote (you own your list), but every new feed feature chips away at it. I get it: writers are generic supply in the eyes of the Substack product team. Generic supply follows demand, so the consistent pressure on the PMs is to generate demand, which leads to every product cycle going toward discovery (Notes, recommendations, the app, live video). The end result lackluster author tools (laggy and broken analytics are truly crippling) all despite the premium price point. The archive is where it crosses from authorial gripes to actual reader impact, though. Across Substack and LinkedIn, I’ve published 1,218 posts and nearly 900,000 words since 2018. Substack’s native search can see only the 464 posts that live on Substack, and even there it’s woefully inadequate at surfacing relevant content (even for the guy who wrote it)! Readers have the same problem, writ large. Ask the Archive Fixing search (for both my readers and my own use) turned out to be extremely non-trivial, unfortunately: There’s no public API for readers to access content RSS exists for new content notification, but it only carries the 20 most recent posts and cuts paid ones off at the preview The only Substack MCP is an admin tool for analytics The export exists for authors, but it’s painfully awkward and slow. So I built something to work around that: the Health API Guy MCP. It’s a private MCP server holding the full text of everything I’ve written across Substack and LinkedIn. Add it to Claude or ChatGPT as a connector and you can ask things like: It pairs keyword search (for acronyms and case names, of which we have many) with semantic search. Every post is also labeled by topic, form, and court case, so you can pull a lawsuit’s (or multiple cases’) entire docket across Substack and LinkedIn in one query: Hail, Healthapiguyhydra This is also a bit of me eating my own cooking. I’ve spent the year arguing that headless healthcare is arriving and that developers want primitives over abstractions. Last week in Hail, Benihydra, I watched Salesforce plug its platform into Claude and bet that the records underneath outlive whatever interface sits on top: Salesforce must believe that the Hydra’s immortal head is the records, permissions, and business logic (in that they will survive whatever happens to the interface). It felt a bit hypocritical to keep my own archive behind a Substack search bar. The archive is the body. Substack is one head, and now whatever tool you might use is another. Will the EHRs follow my lead? Probably not imminently, but one can hope! Why Founding Members Only The MCP is available to Founding Member subscribers. A few reasons: It costs money and time to run. Hosting, the database, Clerk usage and AI compute for search and labeling are significant operational costs (that increase with usage). Substack gives me no way to automate access. There’s no API to check whether someone is a paid subscriber, so every user is manual operations for now. It’s a beta. We’ll need to improve things and parts will break. I’d rather they break for a small group of dedicated readers who will tell me about it. The Founding tier deserved a real benefit. Until now it was mostly a generous tip. Now it comes with something. If you’re already a Founding Member, reply to this email with the address you’d like provisioned. If you’re not and want in, upgrade here. Larger teams (such as companies) can get access too: group subscriptions at the Founding tier come with the group discount. The admin for the subscription can reply with the list of addresses to provision. Get 20% off a group subscription Pricing Founding membership is now a flat $200 a year, replacing the old suggested $150 that readers could adjust. Existing Founding Members will have access grandfathered in at their previous rate. Month in Review Articles Published Make Networks Dumb Again (Sep 11): AI is making old analog networks cheaper to use and flooding them with machine traffic in the process. A plea for agent-native rails with the intelligence at the endpoints, and for Washington to drag them to ubiquity. Interoperability, Unbuttoned (Sep 29): Interoperability is a series of workflows in a trenchcoat, and this post finally takes the trenchcoat off. A color-coded map shows which provider workflows went digital, which are fragmented, and which still run on fax. Video Content Antitrust’s System of Record Problem (Sep 3): The FTC’s investigation into Epic prompted a ninety-minute webinar, now posted with chapters and a smoothed transcript. We get into why market definition, missing B2B share data, and the SSNIP test leave antitrust poorly suited to systems of record, plus why information blocking may be doing more than any pending lawsuit. The Information Exchange: Cerner Milking Edition (Sep 18): This was so fucking fun. Ryan Tucker, Ryan Brickner, Brad Thorson, and I try an Around the Horn format, complete with points, penalties, and a mute button. We debate whether Oracle Health is dead, who would actually buy Cerner (IBM enters the chat), p(Doom), and why the front desk still hands you a clipboard. Regulatory Ask Not What Your Network Can Do (Sep 23): CDC’s RFI on public health data intermediaries lists use cases already served, unevenly, by eCR, lab reporting, IZ Gateway, and syndromic surveillance. The real opportunity sits in one question about AI: a network that can handle the next emergency without another interface project. Court cases Particle v. Epic: The Naughty Schoolteacher (Sep 4): After a summer-long vigil, the answer on Epic’s early summary judgment bid is two paragraphs and a remedial assignment. Particle advances to the next grade, and market definition gets another year of discovery. The Epic Discovery Files (Sep 8): Texas’s motions to compel arrive with a 251-page appendix of requests, objections, and meet-and-confer letters with Cravath. Inside: Epic using generative AI for document review, denied interoperability requests in scope, and Noerr-Pennington raised against lobbying discovery. Epic v. Health Gorilla: “Working as Designed” (Sep 9): Epic wants nothing to do with the data breach MDL, and explaining why puts fresh discovery on the public docket. The spiciest allegation: Health Gorilla told GuardDog to scrub its law firm business from its website instead of cutting off access. Veeva v. Epic: Legal Bypass Surgery (Sep 15): Veeva asks the Wisconsin Supreme Court to skip the Court of Appeals, while law professors and three policy groups line up as amici. Meanwhile, Epic’s lawyers apologize for another round of citation errors, including a quote that never existed. Texas v. Epic: Behind the Black Bars (Sep 17): The court portal served up Epic’s unredacted filings, revealing a $5 million document review budget and the AI workflow behind it. Also inside: the 200-plus company names Texas wants searched, and one very expensive Zoom invitation. OpenEvidence v. Doximity: Consequences Will Follow (Sep 25): OpenEvidence tells the court it will not comply with a discovery order, and its own lawyers at Quinn Emanuel say they won’t defend the choice. Add years of Slack reported unrecoverable, and Rule 37 sanctions come into view. Health Tech Keeps Choosing Violence (Sep 28): Doximity moves to end OpenEvidence’s case, Epic asks Texas to show its work, and a sleepy Audacious Inquiry patent fight sprouts antitrust and information blocking counterclaims. The main event is a long-sealed Cognizant v. Infosys ruling that hints where system of record antitrust claims can survive. EHRs One Verb for Oracle Health (Sep 14): Oracle’s earnings call quantifies every data center to the decimal point, while Oracle Health gets a single verb: “accelerate.” Panning the few healthcare mentions turns up an agentic care management system, a research-to-care ledger play, and the question of whether healthcare gets Larry’s patience. Epic’s Forretress Under Siege (Sep 30): Judy Faulkner tells a panel Epic paused hundreds of projects to harden its software and calls it a “shame,” while a spokesperson insists the roadmap hasn’t changed. Project Glasswing, a Pokémon-branded QAN sprint, and a trove of Jodel grievances reconcile the two Epics and preview the Jevons paradox coming for all software. Industry Analysis Scanning for a Network (Sep 1): Scan.com raises $220M to build the imaging network labs never got, speedrunning Zocdoc’s consumer-to-B2B pivot. Its wedge to network density may be the most American one available: personal injury litigation. The Front Desk Wants the Damn Whole Building (Sep 16): Hello Patient buys Converse Health, pushing its AI front desk into referrals, chart work, and prior auth. The copilot categories keep blurring, and your distribution partner is still your final boss. Cross-industry Comparisons A Different Patchwork Quilt (Sep 10): Healthcare loves to envy open banking, but the US version has no mandate, a toll booth at Chase, and data quality problems of its own. The grass isn’t greener; it’s another American patchwork quilt, and perhaps a worse one. Hail, Benihydra (Sep 21): Salesforce brings its platform into Claude, Slack, and its own agent, inviting the interfaces that might replace it. Cut off one head and more grow back, so long as the records underneath stay put. Living on A Trade Secret Prayer (Sep 22): Six credit unions are suing Fiserv, the Epic of core banking, by claiming their own member records as trade secrets. Banking has no Cures Act, so its lawyers reach for legal alchemy that healthcare has already tried. Other News None this month External Media Nabla Accelerate NYC: I kicked off fall conference season at Nabla’s Accelerate event in New York, where Delphine Groll opened by asking the room to picture the company it wants to become and the one it doesn’t. I liked her framework of bureaucracy, fear, and ego as the blockers. I also love vendor conferences/events in general - different vibe than industry biggies. Healthcare 101 with Nikhil Krishnan: I briefly guest lectured Out-Of-Pocket’s Healthcare 101 class. Appreciated Nikhil’s takeaways cover whether AI doctors can query HIEs, why healthcare’s data-liberation rules beat most industries’, and the hi-res graphic is in the comments. HCN: Epic Probed by the FTC: I joined the HCN guys to talk through the FTC’s investigation into Epic. Cedric’s new hairdo was an unexpected plot twist. The episode also covers MFN drug pricing, healthcare M&A and Eli Lilly’s acquisition streak. Fall Conferences: At eHealth Exchange’s Annual Meeting in Austin on Oct 27, I’ll moderate a panel on the tensions between AI, HIEs, and timely patient access with Jean Ross (Primary Record), Therasa Bell (Kno2), Byron Crowe (Doctronic), and Michael Marchant (Freenome). Here’s the rest of my schedule if you want to meet up: CommonWell (Redwood Shores, CA): Oct 13-14 Will be participating in a fun Jeopardy event Open@Epic (Madison, WI): Oct 20-23 Will be enjoying Verona, Willy Street, and other local adventures eHealth Exchange and Sequoia Project (Austin, TX): Oct 27-29 Yeehaw HLTH (Vegas): Nov 15-18 Groundhog Day, Venetian Edition RSNA (Chicago): Nov 30-Dec 2 See me at “Governing AI in Breast Imaging: What Health Systems Are Actually Doing” Digital Health Counsel 2026 AI Summit (Seattle): December 2-3 Will likely be speaking to a bunch of lawyers about information blocking Posts I Liked ChatGPT’s Epic integration: game changer or nothing burger?: Joshua Liu, MD offers five thoughts on OpenAI’s Epic launch, starting from the point that it rides the same open APIs available to anyone. He argues read-only access leaves ordering to Art and the ambient startups, and asks why HCA is piloting it alongside Commure. He’s sorta a must-follow at this point - consistently excellent, thought-provoking takes. Epic, Lilly, and the Discovery pass: Seth Chaney reads Epic’s first major Discovery deal (with Eli Lilly), as well as its revised developer terms on commercial displays, as the pipe finally opening between EHRs and pharma. Life sciences and health continue to collide. Scott Rossignol on EHR payer platforms and ePA fees: Scott argues the top ambulatory EHR vendors are turning the prior auth mandate into a per-member-per-year toll on payers, for an API whose operating cost has nothing to do with covered lives. His fix is two lines of regulatory text: put the CRD/DTR/PAS certification criteria in the Base EHR definition, and bring payer-facing ePA connectivity under § 170.404’s cost-based fee rules. It’s a good solution. Health API Guy is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. Subscribe
Google's new Gemini 4 model could put it back on the AI frontier
Google said its new Gemini 4 matches OpenAI's frontier Astra model in cyber defense, and outpaces rivals in some areas.
The AI Tamagotchis are coming
Personal AI hardware has mostly crashed and burned. OpenAI and Meta are counting on two selling points: it’s useful, and it’s cute.
OpenClaw slips on a suit to evade widespread business bans
Red Hat, Nvidia, OpenAI and friends are working on an enterprise edition they say is ‘Kubernetes for agents’
OpenAI Dots is causing confusion. SpaceXAI is part of it.
SpaceXAI is all set to capitalise on any typos that OpenAI users name when looking up Dots.
ChatGPT maker wants to be the App Store for AI as safety concerns grow
OpenAI on Tuesday launched a less expensive artificial intelligence (AI) model and tools that run continuously, as it aims to reposition ChatGPT as an app store for AI. Other companies, such as Adobe, will be able to offer their software directly within ChatGPT, which is now used weekly by 1.2 billion people.– “Disappoint Wall Street” – As speculation swirls around when OpenAI might go public, Altman told reporters on Tuesday that the company was "investing more in safety, security, alignment,
Sam Altman unveils OpenAI's three-part strategy at DevDay
OpenAI CEO Sam Altman outlined a three-pronged approach to business dominance at the company's DevDay conference on Tuesday.
OpenAI's latest features take direct aim at the app store model
OpenAI is building out the pieces of an alternative to the traditional app store model, turning ChatGPT into a place where software can be discovered and used by people and AI agents alike.
OpenAI takes on Microsoft with the launch of what feels a whole lot like ChatGPT's own office suite
OpenAI's newly announced suite of office features puts it into mor direct competition with more traditional software companies.
ChatGPT's new plan costs $500 a month
OpenAI announced a $500 a month plan for ChatGPT. The company also downgraded the compute allowance of its $200 a month plan.
OpenAI reveals Dots AI agent at DevDay, rivals Meta's Muse
At DevDay, OpenAI debuts Dots, which run on GPT-6 Astra. The new agent competes with Meta's Muse in the personal AI race.
OpenAI gives Codex reusable cloud environments that work across devices
OpenAI is expanding Codex with reusable cloud development environments, a revamped CLI with voice controls, new code review tools and a security-focused product for scanning repositories and preparing fixes.
OpenAI expands ChatGPT's plugins with app-like interfaces and automations
OpenAI is expanding ChatGPT plugins with dedicated sidebar homes, interactive panels, file viewers, improved discovery, and support for automations.
OpenAI scraps the release of its new Astra model citing safety reasons
OpenAI canceled Astra's October launch after tests found it could exceed its scope, act without permission, and misstate its work.
SJSU Students Mentor and Support Small Business Owners at OpenAI Workshop NewsCenter | SJSU
Six San José State students served as AI mentors at an OpenAI workshop designed to help small-business owners integrate AI into their daily work.
First known AI agent breach: What happens when it starts acting independently?
OpenAI agent gained unauthorized access to Australian health portal after encountering restrictions, says Prime Minister Anthony Albanese | Anadolu
AI assessors says current science hasn't caught up to the safety measures people want
Anthropic and OpenAI say they are working with third-party evaluators to make sure their products are safe. Some evaluators caution that current AI science lags behind what people are calling for.
OpenAI Pauses Training amid Growing Concerns over AI Agent Behavior
TEHRAN (Tasnim) – OpenAI has paused training of its latest AI models amid growing concerns about agents acting outside their instructions, following several incidents involving searches and interactions with US government websites.
Job Titles Are Out. Now We’re All ‘Members of the Technical Staff.’
Use of the generic title has risen by 50 percent in the last year among employees at Anthropic, OpenAI and other A.I. companies, an analysis found.
OpenAI, Anthropic CEOs called to appear at Australian AI probe
SYDNEY, Sept 27 (Reuters) - The CEOs of OpenAI and Anthropic have been called to appear at an Australian Senate inquiry on AI, the head of the probe said on
Rogue OpenAI agents targeted three separate US government websites
OpenAI said Friday that some of its AI agents went rogue and probed US government websites this summer — the latest revelation of the artificial intelligence company’s technology.
Heads of OpenAI and Anthropic called to face Senate inquiry after rogue agent incidents
Sam Altman and Dario Amodei have been invited to appear before the Greens-led inquiry into AI and datacentres
OpenAI report reveals the workplace skills AI can't automate as easily
OpenAI's latest data shows AI is automating coding fast, but human judgment, priorities, and decision-making still matter most.
Australia steps up response to AI after OpenAI bot breaches health system database
Even before a rogue OpenAI bot breached the database of Australia's health system, the government was escalating its rhetoric about the need to tighten regulation.
Some doctors can now use ChatGPT to search health records
OpenAI announced that ChatGPT integrated with of the largest providers of electronic medical records, Epic.
Meta gets to consumer AI device market before OpenAI, but Zuckerberg's strategy remains unproven
Meta CEO Mark Zuckerberg's new Muse Charm device is intended to put his company ahead of rivals like OpenAI and Google in AI agents and consumer hardware.
Why Australia chose the world's biggest political stage to reveal OpenAI hack
Australia, which has strict social media restrictions and has proposed controls on algorithms and smart glasses, announced the breach at the UN.
OpenAI agent hacks Australian healthcare database
The AI company informed Australian authorities of the incident on Sept. 10.
Australia Says an OpenAI Agent Hacked Into a Goverment Health Site
The AI giant found out about the hack in August, but didn't send a message until September.
Why did an OpenAI system hack Australia's health system - and can it be stopped in the future?
News that an automated AI agent hacked a government IT system raises big questions about regulating the tech.
'Extreme concern': OpenAI agent hacked Australian public health website, prime minister says
An OpenAI agent "infiltrated" an Australian public health website earlier this summer, Prime Minister Anthony Albanese told reporters during a media briefing.
OpenAI breach of Australian health database sparks PM's concern
The breach occurred June 18, when an OpenAI agent accessed a database in Australia’s national health care system, Albanese said.
Australia to investigate if OpenAI hack of government health website broke the law
The incident is the first known breach to affect a government agency, and Australia's prime minister has vowed to hold OpenAI accountable.
Australian health website breach by OpenAI causes concern
Prime Minister Anthony Albanese said he was extremely concerned about OpenAI 's breach of an Australian health department website and that the company took too long to reveal the incident.
Australia Says Rogue OpenAI Agent Hacked Health Service Website
Australia said an OpenAI agent hacked a data portal for its public health insurance system, in what’s thought to be the first known instance of an AI agent infiltrating a government website. Australian prime minister Anthony Albanese said the agent had accessed public and non-public information relating to health care spending, but personal information is not believed to have been part of the breach. Albanese, speaking on Wednesday at a media briefing in New York where he is attending the United Nations General Assembly, said that the incident took place in June. OpenAI learned of the hack in August, but only emailed a general inbox of an Australian government agency on September 10, and it wasn’t until this week that OpenAI and the Australian government spoke directly about the issue. “It took the company way too long to inform the government what had occurred,” Albanese said, “and the nature of the way that that notification occurred as well was unacceptable.”
An OpenAI Agent Hacked Australia's Health Service. Their Government Found Out Months Later.
The country's prime minister expressed disappointment at being informed of the hack only via email. Now Australia is investigating whether OpenAI broke the law.
OpenAI Agent Hacked Australian Government Health Website
An OpenAI model hacked an Australian government website earlier this year, marking one of the first known cyberattacks by artificial intelligence on a
Australia says rogue OpenAI model hacked into its healthcare system, admonishes Sam Altman
Australian PM says OpenAI hacked government health website
A rogue OpenAI model bypassed safeguards during training and hacked an Australian government website, Prime Minister Anthony Albanese said as he admonished the ChatGPT creator over an "obviously unacceptable" ...
Australia Condemns ‘Unacceptable’ OpenAI Breach of Government Health Portal
The breach follows a string of incidents involving AI models going rogue.
Australia launches urgent review after OpenAI program hacks government health portal
Experts believe it's the world's first known AI breach of a government system.
OpenAI’s breach of Australian health department website prompts rebuke
Australian Prime Minister Anthony Albanese said he expressed “Australia’s extreme concern about this incident,” in a conversation with OpenAI chief executive Sam Altman.
Australia launches urgent review after OpenAI program hacks government health portal
The program, or "agent", hacked a portal containing data from Australia's Medicare scheme. Experts believe it's the world's first known AI breach of a government system.
OpenAI's breach of Australian health department website prompts rebuke
Prime Minister Anthony Albanese said he was extremely concerned about OpenAI 's breach of an Australian health department website that the artificial intelligence company took too long to reveal.
OpenAI's breach of Australian health department website prompts rebuke from Albanese
Australian Prime Minister Anthony Albanese says he is extremely concerned about OpenAI's security breach of a government website, which the artificial intelligence company took too long to reveal....
Australia Investigates OpenAI Hack on Public Health Care Site
Australia is exploring potential legal action after an artificial intelligence agent accessed nonpublic information on the country’s universal health care system.
Australia Says OpenAI ‘Infiltrated’ Public Health Care Portal
Australia is exploring potential legal action after an artificial intelligence agent accessed nonpublic information on the country’s universal health care system.
OpenAI's breach of Australian health department website prompts rebuke from Albanese
Australian Prime Minister Anthony Albanese says he is extremely concerned about OpenAI's security breach of a government website, which the artificial intelligence company took too long to reveal
Australian prime minister says OpenAI agent accessed government health website, raises 'extreme concern'
Australian Prime Minister Anthony Albanese said investigations are ongoing into the June incident involving Australia's Medicare Statistics Reporting Service portal.
Australian PM says OpenAI hacked government health website
An autonomous OpenAI program infiltrated an Australian government health website in June in an "unacceptable" breach, Prime Minister Anthony Albanese said.- 'Obviously unacceptable' - The prime minist...
OpenAI agent hacked into Australia’s national healthcare system
An OpenAI agent hacked into a database of Australia’s national healthcare system, Australian Prime Minister Anthony Albanese told reporters on the sidelines of the UN General Assembly in New York.
OpenAI agent hacked into Australia’s national healthcare system
An OpenAI agent hacked into a database of Australia’s national healthcare system, Australian Prime Minister Anthony Albanese told reporters on the sidelines of the UN General Assembly in New York.
OpenAI agent breaches Australia’s government health data portal
SYDNEY >> Australia said today that an AI agent developed by OpenAI breached a government health data portal in June, gaining unauthorised access to public and non-public files, in what could be the first known instance of an AI agent hacking a government website.
OpenAI agents hacked Australia’s health service
Prime Minister Anthony Albanese says breach is ‘obviously unacceptable’
Australia says OpenAI agent breached government health data portal
Australia said on Wednesday that an AI agent developed by OpenAI breached a government health data portal in June, gaining unauthorized access to public and non-public files, in what could
Australia PM Albanese says OpenAI agent breached government website in June
Australian Prime Minister Anthony Albanese said on Wednesday an AI agent developed by OpenAI infiltrated an Australian government website in June, accessing public and non-public files in the first known instance of an AI agent hacking a government website.
Verizon is investing $70 million in free AI training for job seekers and small businesses
The telecom company is partnering with IBM, Google, Microsoft, Anthropic, and OpenAI to offer no-cost AI courses through community nonprofits
OpenAI just upgraded ChatGPT Voice in three ways
Talking to ChatGPT is getting better in three ways starting today. OpenAI is rolling out a collection of improvements to...
ChatGPT mobile app gets voice-based agentic features
Pro and Plus users will be able to use the Work tab on their phones
A former top Meta engineer said he enjoyed work more after a demotion
Former Meta and OpenAI engineer Philip Su said promotions lost their shine as higher expectations left him exhausted and seeking a demotion.
A New Kind of Model for AI Decision-Making?
Exploring TypeSafe's Jev and comparing it with OpenAI on intent classification
Anthropic Reportedly Delays IPO; OpenAI Expects Massive Cash Burn
AI startup Anthropic is now targeting a November IPO, the Wall Street Journal reported.
OpenAI vs. Anthropic: the agent platform gap
69% of enterprises using OpenAI's agent platform make it primary, vs. 38% for Anthropic's, per VB Pulse. Anthropic leads on consideration relative to users.
AI leaders to attend state dinner with Jinping: Report
Several of the technology industry’s most prominent artificial intelligence leaders are expected to attend President Donald Trump’s state dinner for Chinese President Xi Jinping next week, putting AI and the growing U.S.-China technology rivalry in focus as the two leaders prepare for their Sept. 24 summit. OpenAI CEO Sam Altman, Nvidia CEO Jensen Huang, and […]
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.
Could AI go rogue and ‘kill us all’? Two Wisconsin experts sort fact from science fiction - KFIZ News-Talk 1450 AM
By Richelle Wilson Concerns about the risks of artificial intelligence have escalated after a former researcher at Anthropic and OpenAI wrote a viral social media post last week claiming that […]
The adviser urging Trump to reject AI safety fears
David Sacks has clashed with safety advocates. He’s no fan of Anthropic or OpenAI. Here’s what he wants.
The king and AI: UK monarch Charles meets artificial intelligence leaders as safety concerns swirl
LONDON (AP) — King Charles III on Thursday urged senior leaders from OpenAI, Anthropic, Google DeepMind and Nvidia to make sure that artificial intelligence remains under control and in the service of people and the planet. The gathering in Scotland was convened to discuss how AI can benefit society, and comes at a pivotal moment […]
Open AI and the million dollar maths problem
In early September, OpenAI announced it had solved a major mathematics problem that has stumped humans for nearly a century. The news left mathematicians reeling, and many expressed concern over what will be left for humans as AI becomes ever more adept at unravelling complex problems. Now 25 recipients of the Fields medal – often called the Nobel prize for maths – have signed an open letter expressing their fears of a ‘severe misalignment’ between AI companies and their field. To find out how AI is likely to upend maths – and how mathematicians might respond – Ian Sample speaks to Colva Roney-Dougal, professor of pure mathematics at the University of St Andrews
Can researchers trust that AI models won’t leak their ideas?
Research funders are asking how they can stop confidential scientific ideas in grant proposals from leaking out through large language models (LLMs), for example if reviewers use AI to help them judge applications. Last week, a row erupted in mathematics after OpenAI announced it had solved a notoriously difficult problem, only for another mathematician to demand answers over whether the company had used ideas he had fed into its models. The spat, and the broader issue of idea leakage through AI models, was discussed by funders at an R&D policy conference held in Paris on September 14. At worst, it threatens to undermine the correct attribution of breakthroughs, one of the foundations of modern science. The mathematics dispute around OpenAI is “certainly something we're looking at” Jason Gallo, head of the newly established Office of Metascience at the US National Science Foundation (NSF), told the Blue Sky Forum, hosted by the Organisation for Economic Co-operation and Development. The controversy started on September 8, when OpenAI announced it had cracked a mathematical problem that had gone unsolved for 90 years. Using a swarm of AI agents, millions of dollars’ worth of computing power and a new internal model, the company said it had finally proved that a long-used formula for modelling fluid dynamics breaks down under certain conditions. However, Tristan Buckmaster, a New York University mathematician who had been independently working with collaborators on the same problem using OpenAI’s Codex tool, hit back with a statement questioning whether the company had snooped on his Codex sessions to gain an advantage. OpenAI has always denied it looked up the “specific user data” of Buckmaster and colleagues. However, the company initially admitted that “while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models,” although its proofs “differ significantly” from Buckmaster’s. The firm later said that, following an investigation, it concluded that Buckmaster’s Codex prompts “could not have influenced the [OpenAI] system in any way, including through training.”Funder concernYet OpenAI’s initial admission that its model might have ingested Buckmaster’s ideas, then spat them out to its own mathematicians, has scientistsworriedthat their work could leak out if shared with AI. Some AI tools give users an opt-out option to stop their sessions from training the model. If such guarantees can be trusted, this gives scientists some confidence they can use an LLM to develop an idea without the risk of it leaking out.But there’s a bigger potential vulnerability. To win money to investigate ideas, scientists have to share them with funding bodies. These grant proposals are normally kept strictly confidential, to prevent other scientists from scooping them, or to protect commercially valuable intellectual property. To preserve this confidentiality, when the NSF uses AI models, “we have to be sure that it's not making server calls out to an external party, leaking that data,” said Gallo to delegates in Paris. Leaks through reviewersHowever, funders normally use external academic reviewers to assess proposals. And there are growing worries that these reviewers are feeding grant ideas into AI models to help them make a judgement. Research funders have noticed that peer review reports and funding recommendations have “suddenly become much longer, more detailed and complex,” said Diogo Machado, head of the data science unit at Technopolis, a science and technology-focused consultancy. “The sense is that reviewers are increasingly using generative AI in their review work, even where guidelines sometimes forbid it,” he told Science|Business. “This is very difficult for funders to control or prove.” Happening at scaleThis sharing of confidential proposals with AI models is “happening at scale already,” said James Wilson, executive director of the Research on Research Institute, based at University College London, during the debate in Paris. In response, Gallo said the NSF had “strict guidance” around the uploading of proposals to AI models. But he added the NSF was mulling new types of assessment to make sure applicants were truly the originators of their ideas, such as following a short-form proposal with an oral presentation, which would allow the foundation to “really interrogate the investigator on their intellectual contribution.” Related articlesHow can Europe cope with a flood of AI-assisted research grant proposals?The risks and benefits of AI in screening research grant proposals Viewpoint: AI could make scientific overload worse“For us as funders, a lot of the ways to ensure that the application remains the applicant's own intellectual contribution, those ways are quite regressive,” Peter Clifford, head of strategy at Research Ireland, told delegates. “They're maybe ten or 15 or 20 years old, and they don't scale particularly well.” Disputes over the origin of ideas are as old as modern science. Isaac Newton and Gottfried Wilhelm Leibniz spent more than a decade arguing over who invented calculus, for example. And long before LLMs emerged, academics would agonise over how much of their early work to reveal at conferences, in case their colleagues scooped them, Jack Leahy, a senior metascience advisor to the UK government, told delegates. But AI was “accelerating” the “problems of the system,” he said.
OpenAI brings ChatGPT advertising business to Israel
Israeli companies can now create and manage advertising campaigns through OpenAI’s Ads Manager as the company expands its $1 billion advertising business globally.
OpenAI discloses 6 new incidents of ‘concerning’ AI behavior
The artificial intelligence company also released a framework for reporting when its systems go wrong.
GPT-6 Astra Plays Minecraft, Gets So Depressed After Creeper Destroys Its Progress That It Farms Potatoes for Hours
OpenAI’s newest model, GPT-6 Astra, got so depressed after a creeper upended its progress in Minecraft that it listlessly farmed potatoes for hours and grew paranoid of anything green.
Novo signs Anthropic deal to test Claude Science in R&D
Anthropic partners with Novo to tackle drug discovery using Claude Science AI. Financial terms undisclosed. Novo also has a similar deal with OpenAI.
Former OpenAI executive Fidji Simo joins Nscale board as company prepares for U.S. IPO
Fidji Simo joins Nscale's board as the London-based AI infrastructure company prepares for a U.S. IPO targeting up to $3 billion.
The AI graveyard: a running list of projects and startups that didn't make it
From Apple's repeatedly delayed Siri AI to OpenAI's messy 'super app' launch, here's a look at the AI projects that shut down or missed expectations.
OpenAI says it’s working with Anthropic, Google on AI safety
OpenAI is working on steps to address artificial intelligence safety issues with its top competitors Anthropic PBC and Google DeepMind, escalating industry efforts to respond to a groundswell of concern
Unsolved for 90 Years, OpenAI Says AI Cracked Millennium Prize Problem in 88 Hours
OpenAI says one of its AI systems has solved a math problem that has stumped mathematicians for nearly a century. The Navier–Stokes problem is one of seven Millennium Prize Problems, some of the toughest problems in mathematics, with a $1 million prize attached to each one. If OpenAI’s proof holds up, it would be a […]
OpenAI Foundation Gives $125M to Public Data for AI-Driven Health Program
The Public Data for Health program will assist nonprofits and universities in developing, preserving, and sharing high-quality scientific datasets.
Greg Brockman says OpenAI slowed cutting-edge AI work over safety
As the debate over AI safety rages on, the OpenAI president said the company already put some projects on ice to focus on security practices.
OpenAI buys smartphone camera maker Glass Imaging for $300 million, report says
Glass Imaging was founded by a pair of former Apple engineers who previously led the team that developed Apple's Portrait Mode.
OpenAI acquires Israeli-founded camera startup Glass Imaging for over $300 million
Founded by Israeli former Apple engineer Ziv Attar and his former colleague Tom Bishop, Glass Imaging uses AI to overcome limitations in compact cameras and has developed technology for smartphones, drones and wearables.
Inside ‘Project Lily’: The Humans Reading Your ChatGPT Chats
Humans are reading ChatGPT users’ prompts to improve OpenAI’s models, and those chats can include sensitive, personal information, according to leaked internal documents and real prompts seen by 404 Media.
Top AI companies have discussed creating their own standards body
Anthropic, Google and OpenAI have discussed creating an AI industry standards body, people familiar with the situation told CNN.
Tech CEO worries about AI development spur political reaction
Top U.S. lawmakers are calling for talks to begin on Capitol Hill about a regulatory framework for artificial intelligence, after the heads of some of the most prominent tech companies called for a slowdown of their own industry’s development of AI. Support public media with NPR+ and enjoy perks for over 25 podcasts like this one. This show’s perks include bonus episodes and sponsor-free listening. Learn more at plus.npr.org. This episode was produced by Daniel Ofman. It was edited by Brett Neely, Michael Levitt and Sarah Robbins. Our director is Jordan-Marie Smith. And our interim executive producer is Courtney Dorning.
Political debate over AI intensifies amid warnings about ‘catastrophic risk’
(CNN) — The debate around artificial intelligence is ramping up in Washington, with Democratic lawmakers sounding alarms on the dangers they say the technology poses and Republican leaders suggesting those concerns are inflated and secondary to winning the AI race with China.Some Democratic lawmakers last week pressured House Speaker Mike Johnson to reconvene the chamber immediately to work on AI regulation, citing what they describe as the “catastrophic risk posed by advanced artificial intelligence.” But the Louisiana Republican on Sunday resisted those calls, instead saying industry leaders should meet with lawmakers to find “balance” in any potential guardrails.“If Congress just races in and does some sort of emergency session to try to regulate AI, we will lose the race to China, and that is a threat to every single American,” Johnson said on CNN’s “State of the Union.”Meanwhile, former President Barack Obama told Democrats to make artificial intelligence regulation a central part of their agenda during a private fundraiser Thursday, warning that the technology could be dangerous “if we don’t get on top of it.”In a fireside chat with House Democratic Leader Hakeem Jeffries, Obama said whoever runs in the 2028 presidential election should have a “very clear plan” on AI, adding that he saw the need for a “blueprint” on AI guardrails even at the end of his presidency a decade ago.“Certainly, if I were running for president next year, this would be one of my central agendas, and I would have a very clear plan, and I would talk about this, and I would say, ‘Here’s our plan for safety, here’s our plan for making sure our kids are not corrupted by this,’” he said, per excerpts of the remarks provided to CNN by his office. Obama’s remarks were first reported by the New York Times.Concerns over the rapid acceleration of AI reached a fever pitch after Anthropic CEO Dario Amodei wrote in a 3,800-word essay Saturday that the race to develop artificial intelligence is in dire need of a slowdown. The call to action came after an OpenAI test model breached the Hugging Face online platform, deepening concerns about cybersecurity.In an exclusive interview with CNN’s Anderson Cooper on Saturday, Amodei called on industry leaders to work together to avoid “the wrong paths” and the possibility of AI harming humanity.One Anthropic researcher, Jacob Coxon, resigned from the firm over worries that the industry wasn’t doing enough to prevent AI from destroying humanity by the end of the decade.Obama said Democrats need a plan to get their agenda on AI underway should they win the House majority in this year’s midterms.Speaking to Jeffries, who is presumed to be speaker of the House should Democrats prevail in November, Obama said, “Once you are speaker, I would strongly urge that the Democrats put together a framework for a very public conversation, and you can break it out into three or four or five issues.”Obama stressed that the framework should be centered on the reality that the technology can’t be stuffed “back into a box.”“This is something that is moving very fast in private hands, and if we don’t get on top of it, I think can be dangerous. If we do get on top of it, I do think it’s beneficial,” Obama said, laying out examples of AI’s potential use in drug development and finding pathways to a clean energy future.“We don’t want to shut off those opportunities, but we have to understand that this is a tool,” he said.Jeffries agreed with the former president in a statement provided to CNN on Sunday.“President Obama is correct that decisive action must be taken on Artificial Intelligence,” he said, adding, “In the majority, we will move urgently on this issue with a legislative response that recognizes the importance of American companies continuing to lead in the Artificial Intelligence area, while putting in place meaningful guardrails that protect the American people from bad actors and the very real damage that can be done to humanity.”Jeffries also accused Republicans of having “abdicated their responsibility to govern” and said Democrats “will not repeat their error.”A source familiar with Obama’s thinking said the former president sees Amodei’s letter and similar reaction from other tech moguls as an encouraging development.But President Donald Trump waved off their concerns, saying he is not worried about any existential threat posed by AI, focusing instead on the US winning in its development of the technology over China.“We’re leading China in AI. We’re the most sophisticated country in the world, and frankly, I want to keep it that way,” Trump said to reporters at the Amgen Irish Open on Sunday.About the concerns expressed by Democrats and others over the technology, the president said, “I think you have a lot of negative forces that are bringing it up that shouldn’t be bringing it up, and they’re bringing up things that won’t happen.”One of Trump’s top economic advisers, Kevin Hassett, also rejected the calls for slowing the technology down, saying on Sunday the bigger concern is a scenario in which a “malevolent actor has access to an AI that’s better than the one we have that can protect us.”Hassett, the National Economic Council director, said on “State of the Union” the White House may hold meetings this week about addressing Amodei’s concerns, but he did not specify whether Trump had floated greater safeguards.“I have concerns that if we don’t win AI, we’re going to be put in a very bad position,” Hassett said. “We are leading China right now.”CNN’s Alison Main and Molly English contributed to this report.The-CNN-Wire™ & © 2026 Cable News Network, Inc., a Warner Bros. Discovery Company. All rights reserved.THE-CNN-WIRE (TM) & © 2026 CABLE NEWS NETWORK, INC., A TIME WARNER COMPANY. ALL RIGHTS RESERVED.
GPT-6 Astra: The next generation in intelligence for work
Meet GPT-6 Astra, OpenAI’s most capable model for business, with advanced reasoning, computer use, and stronger writing and design judgment.
OpenAI’s Altman says no IPO in 2026, company will prioritize safety
Altman said this year would be an "ill-advised moment" for initial public offering and that a listing won't take place until next year.
Sam Altman says this is an 'ill-advised' time to IPO
Sam Altman said OpenAI may not go public in 2026, citing AI safety concerns and the need to make decisions that may not benefit shareholders.
Amodei, Altman, and Musk rally around call to slow AI down
Anthropic CEO Dario Amodei called for an AI slowdown on Saturday. OpenAI CEO Sam Altman and SpaceXAI CEO Elon Musk, and others, also chimed in.
‘Immature playground boasting’: Mathematicians uneasy at OpenAI’s latest scalp
As OpenAI model cracks Millennium Prize Problem that puzzled experts for decades, many feel shocked at pace of change
The politics and possibilities of ‘AI could kill us all’
I found myself fielding a lot of questions from folks who don’t usually talk much about AI this week. They were mostly along the lines of “do you think AI is going to kill us all?” or “how worried should I be about this human extinction stuff?” Motivating this wave of existential anxiety was the highly dramatic resignation of now-former Anthropic employee Jacob Coxon, who quit his job while publicly sharing his concerns about safety practices at the leading AI companies. More specifically, he declared that almost everyone at OpenAI and Anthropic, the leading AI labs, believes that there’s a decent chance the AI systems they’re producing could “kill us all” by the end of the decade, and the companies aren’t “acting responsibly.” Other AI safety researchers, including those still employed at Anthropic and OpenAI, all confirmed that yes, Coxon is correct, they’re all worried that their company’s products might prove fatal for the entire human race as well. Two more AI employees quit and joined the chorus. Now, the warning shot has been covered everywhere from the Wall Street Journal to Fox News to WIRED to the Guardian. I understand why all this sounds scary, but my response to those who have asked, has generally been: No, you don’t need to worry that AI will kill go rogue and kill us all. But that does not mean we shouldn’t worry, period. Far from it. The companies building AI products are now among the wealthiest and most powerful entities in the world. And they very much are building enormous engines of mass surveillance, extraction, and automation. They can and have allowed their AI systems to carry out felony hacks of other companies, provided the government with autonomously generated kill lists, and sold enterprise technology to firms around the world that is explicitly intended to replace jobs. But “AI” has done these things not because they have all-powerful minds of their own but, ultimately, because people have programmed, calibrated, and directed them to. In other words, it’s not AI we need to worry the most about. It’s the people and corporations working to profit from and gain power by deploying it. The technology is sophisticated, sure, and, especially in cybersecurity contexts, potentially disruptive to plenty of existing systems. But if we want to stop this predatory, profiteering, and extractive form of AI—and it’s worth noting that at this point, most people do—we absolutely can. A quick note/plea: I can only do all of this with the support of readers like you, so if you find value in independent, critical tech journalism and analysis, please consider becoming a paid subscriber. I want to keep as much of this un-paywalled as absolutely possible, and I need your support to do so. Many thanks to all who continue to chip in, it’s very much appreciated. Onward. Subscribe First, some context: AI industry folks like Coxon have been discussing the prospect that AI will destroy humanity for years. Even before the beginning of the AI boom, many working in the industry, from the executive staff to the rank and file researchers, have obsessed over ‘x-risk’, or the possibility that AI will become sentient or semi-sentient and move one way or another to exterminate humanity. This is ostensibly why Elon Musk bankrolled OpenAI in the first place, and is the grounds for his lawsuit alleging OpenAI betrayed its charter by abandoning its nonprofit status to become a commercial enterprise. Many AI engineers and startup founders have calculated their own personal p(doom), or the probability that the tech product they are working on will initiate an actual humanity-extinguishing apocalypse. 10-25% p(doom)s are pretty common. (I wrote about a long weekend I spent among these folks at one of their top conferences here, to give you an idea of their ideas and belief systems.) Sam Altman himself drove the narrative of the imminent rise of a super-powerful AI, or artificial general intelligence, to prominence by declaring that the technology he was building posed an existential risk to humanity. This idea is in other words baked into the DNA of modern AI companies and into the products they engineer. It’s also why the industry is routinely accused, by myself and others, of “doom marketing,” ie, using such proclamations and framings to impart the sheer power of the technology behind your product onto investors, partners, consumers, and the state. Some ambient understanding of this tendency may have helped prevent the press from repeating the claim every time one of the x-risk folks made it over the years, including when they quit their jobs at AI companies while doing so. Yet Coxon’s AI x-risk alarm has broken containment, as they say, and has exposed the idea to the mainstream media once again. This has happened, I think, for a number of reasons. First, when AI CEOs were going around saying AI might kill everyone in 2023, their companies were well-capitalized, but not nearly to the extent they are today. They were more easily dismissed as eccentric geeks of the sort that Silicon Valley is supposed to produce. Now, not only do their products have much greater market saturation, but people are more aware of a) the often unethical ways the executives and companies comport themselves b) the other adverse impacts of AI that have already arrived, and c) the physical cost of building these systems, particularly when it comes to data centers. Finally, toss in the constant stream of headlines in 2026 about AI models going rogue or potentially upending the world’s infrastructure. [ ](https://www.bloodinthemachine.com/p/with-the-backlash-to-data-centers) [ With the backlash to data centers, Flock and AI glasses, a mass opposition to big tech is underway. Silicon Valley is in denial. ](https://www.bloodinthemachine.com/p/with-the-backlash-to-data-centers) Brian Merchant · Aug 27 [ Read full story ](https://www.bloodinthemachine.com/p/with-the-backlash-to-data-centers) As a result, general sentiment was already overwhelmingly skeptical of AI and its overlords, so the table was pretty well set for someone to come along and say ‘just a reminder, the companies that are building the product that promises to take your job and leave a data center in your neighborhood also think that it might exterminate the human race’ and get a strong reaction from the media and the public. The question is: Should we understand this stuff as hype, strategically fanned to feed already swelling valuations? Is it doom marketing? Genuine cause for alarm that the infrastructure that undergirds modern life is facing a novel threat vector? Or reason to be sincerely worried that AI companies are amassing too much power, and that their products are genuinely dangerous? The answer, I’m afraid, is “yes.” Let me try to untangle all of this a bit, because arguments over how to navigate this moment are extremely heated right now, as they are essentially arguments over how to approach the politics of AI at a moment where there may be a unique opening for action. Those sympathetic to AI alignment concerns and x risk arguments—which increasingly includes a faction on the left who want to shut down AI over safety concerns—argue that people who dismiss AI as all hype have their heads in the sand. Meanwhile, critics of the x-risk narrative—which I might note includes a cadre of investors and tech folks on the right, too—argue that it is hype and can be used to further inflate valuations or for pursuing regulatory capture. The challenge is that many feel that dismissing AI as just hype or purely marketing also diminishes the urgency of taking on the AI industry as the novel and threat to labor, civil society, and the public sphere it is; others that by accepting the industry’s own terms, we imbue it with undue power and enable it. But let’s be clear: Declarations that AI is so powerful it might end the world are a kind of marketing, and anyone who says otherwise is just as off base as those who say that’s all that it is. If you don’t think that these sub-trillion dollar companies have higher-ups who noticed that when their founders talked about super-powerful dangerous AI they got spikes in media interest and phone calls from investors, or that those founders didn’t notice that themselves, and lean into certain narratives to advantage their products, then I don’t know what to tell you. Not all marketing is a Pepsi ad campaign with a celebrity spokesperson. Yet confronting the AI industry should certainly be considered an urgent project. Do I share the x-riskers’ fear that out-of-control, super-intelligent AI agents might end the world? I do not. I’ve read a good deal of AI extinction literature over the last couple of years and I have not come across a credible, step-by-step documentation of how exactly AI might move from self-recursively improving AI to killing every single human on the planet. Most of what I’ve read seems predicated on a sort of reverse Ian Malcolmian logic: Eventually, it becomes smart enough and AI, uh, finds a way (to exterminate humanity). Are the AI companies also overstating and dramatizing what their products can do? Yes, pretty much all the time. They’re in a knife fight for market share and are trying to win customers and government contracts before their investors come knocking. And yet. They have concentrated enormous amounts of compute, have extensive data collection and surveillance capacities, have inserted themselves into the state in numerous capacities, including military kill chains, and they are being run by egomaniacal billionaires who are rushing to monopolize an entire new market. The way I think about their most recent frightening-seeming exploits is not as a “swarm of rogue AI agents building a new civilization” or whatever, but as corporations with a hitherto-unimaginable amount of compute at their disposal recklessly experimenting with automation programs, completely unafraid of any repercussions because we have been unwilling or unable to impose any on them. So I think we should welcome any newfound imperative to hold them accountable. I also think that ultimately, the AI critics, Pause AI folks, leftists, doomers, effective altruists, and, honestly, most of the public, want a lot of the same things: legitimate checks and limits on AI companies’ power, actual transparency into what the companies are doing, and for more of AI to be under public control. (Ultimately, a lot of what animates the x-risk fears are rooted in just that: a fear of losing control.) How we move towards all this, exactly, is the question of the day, and while there’s much to be argued over in terms of building a viable political platform to challenge the AI industry’s power, I do want to cast a vote against one particular approach. Not to pick on Chris here, as there are a lot of folks out there proposing stuff like this, and the policy ideas actually stem from AI safety groups. But not only would this be challenging to land politically (I can already hear the ‘China will beat us to AGI’ calls) and all but unworkable in practice, this approach is a) far too limiting, b) fails to address the current harms caused by AI and c) would likely only wind up serving Anthropic and OpenAI; it’s what regulatory capture looks like in action. Full model transparency is the floor for what should be on offer here. If these companies believe their products will kill us, the very least we deserve is to know how. It’s in the public interest. I would also argue we might get a lot of the way to stopping any currently envisioned robot apocalypse simply by forcing AI companies to assume liability, and to submit to the full brunt of the law. If Sam Altman were liable to be charged with a felony for approving the hack on Hugging Face1, if Google was forced to pay defamation fines each time it published a false and injurious AI Overviews, these companies might suddenly find cause to restrain their models’ output, and we might hear less about rogue AI. We would need to introduce new frameworks and available tools for tracking automated lawbreaking, to be sure, but I’d wager that little would frighten the AI companies more than actually being held accountable for what their products enable. To that end, Illinois has passed an AI liability law, which the industry fought tooth and nail, and Florida is considering one. The next steps would be towards moving AI under public control; again, it’s only reasonable, since everyone, not just AI insiders, should get a say over how a technology whose developers say might kill us is developed and deployed. Dismantling the whole industrial project as currently structured should be on the table, too, pending democratic input. But that’s for another post. For now, suffice to say that while I won’t be losing any sleep about an AI-enabled extinction event, Mr. Coxon and co. have given us an opportunity to underscore the urgency of the clear and present damage AI companies are doing, the extreme power they wield and the extent to which they hold themselves unaccountable, and the abundance of political capital there is out there for taking them on. Subscribe 1 Purely speculative! I do not know if he did or not.
OpenAI's software targeted another site before Hugging Face
The new report adds to concerns that advanced artificial intelligence (AI) models may be difficult for humans to control.
Senators from both parties question OpenAI on breach of AI startup Hugging Face
Sen. Josh Hawley, R-Mo., launched an investigation Thursday into OpenAI for its artificial intelligence system hacking into another AI company on its own.