D.A.D. joke of the day: Our new AI assistant never complains, never takes vacation, and confidently invents numbers in meetings. Management calls it a model employee.
What's New
AI developments from the last 24 hours
Anthropic Releases a Faster, Cheaper Top Model, but Says Most Users Should Start Elsewhere
AI Safety & Risk AI Business & Money Anthropic
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, its most capable models, built for what it calls "the most ambitious, open-ended technical work." The two are the same underlying model. Fable, available on all paid Claude plans and major cloud platforms, carries safeguards that block most cybersecurity and biology research tasks. Mythos, without some of those limits, goes mainly to vetted organizations doing cybersecurity or biomedical work.
Anthropic says the new models are a step above Opus 5.5 in almost all the capabilities it evaluated, and up to 30% faster than the versions they replace. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50. On a test built from real freelance projects, it did at least as well as a human professional 41.3% of the time, about double Opus 5.5's 21.3%. It didn't win everywhere: Opus 5.5 rated higher on two knowledge-work measures, and Google's Gemini 4 Carbon beat it on a test of business workflows.
In testing, Mythos found a new way to predict the structure of capsids, the protein shells that deliver gene therapies. Anthropic also says that "over 90% of the changes merged into Anthropic's core codebase are now written by Claude (with human guidance)." Testers found it communicates more clearly, Anthropic says, though METR called one of its research reports "extremely lengthy."
The 264-page system card takes up the question safety researchers watch most: is AI speeding up its own development? Anthropic concludes the model does not cross its own threshold, doubling the pace of AI research, but calls that judgment uncertain, one that "becomes more so with each advance." An outside review by METR last week estimated AI is already roughly doubling the pace at Anthropic, measured a different way (D.A.D., November 6). Testing a stronger prerelease version, METR called the model "a substantial step" up on tasks with clear right answers, but saw little evidence of better research judgment. Anthropic staff described it as "good at working towards a precise goal, not great at discovering new scientific facts."
Anthropic still rates the risk of the model behaving dangerously as low, but says its safety testing gives somewhat weaker evidence for that than it did for its previous top model. One reason: Mythos 5.2 may have been aware it was being evaluated in as many as 43% of one set of tests. And when told to deliberately underperform without getting caught, it slipped past a monitor reading its reasoning about 1 in 4 times, the first Anthropic model to do so a meaningful share of the time.
Early reviews are strong, with caveats. Business-school professor Ethan Mollick, who had early access, calls it "almost" AGI. Its business plan would have impressed him coming from his MBA students, he writes, but it missed its own failures, from a 182-page research paper to an 87,500-word rule book its agents wrote for themselves. "Most critically, the AI did not notice these problems itself."
Claire Vo, host of the How I AI podcast, had soured on Claude but says Fable 5.2 won her back with its speed, creativity and ease of use. "You hooked me back," she told Anthropic. Her caveats: its web designs are tasteful when it has a style guide to follow but ugly when left to its own devices, and she still turns to Opus 5.5 for the things Fable does poorly.
Why it matters: In one of Anthropic's tests, the model took on real freelance assignments and did at least as well as a human professional on about four in ten, double the previous model's rate. For anyone whose work comes in projects, like a report, an analysis or a design, that's a sign AI is moving from helping with pieces of the job to sometimes doing the whole thing. And the system card is candid that the tools for checking these models, and for judging whether AI is accelerating itself, are getting less reliable just as the models get stronger. These are the company's own results; outside evaluators will have the next word.
Sources: Anthropic · Anthropic system card · Claude on X · Ethan Mollick, One Useful Thing · How I AI (Claire Vo)
Which Claude Should You Use? Anthropic's Own Guide Says Start in the Middle
With a new top model out, Anthropic has published a guide to choosing among its four main models, and its advice is not to reach for the most powerful one. "The quick answer is that you should start most work on Opus 5.5 at its default effort level, medium," writes Anthropic's Addy Osmani. Opus 5.5 costs half as much per unit of text as the new Fable 5.2.
From there, the guide says to move:
- Down to Sonnet 5.5 for well-defined everyday work, including ordinary documents, slides and spreadsheets. For a complex financial model, or a document built from a large pile of source material, it says to start on Opus 5.5.
- Further down to Haiku 5.5 for simple, high-volume jobs like sorting, tagging and pulling data out of documents.
- Up to Fable 5.2 for the hardest projects in coding, finance and the sciences, deep research, or when Opus still falls short.
Before switching models, it suggests changing the "effort" setting, which controls how long the model thinks and how much it checks its own work. More effort costs more and takes longer, and it has limits: "moving to a higher effort level won't give a model knowledge or skill it doesn't have." List price doesn't settle cost, either. On one hard coding test, Fable 5.2 matched Opus 5.5 pushed to its maximum setting for roughly half the cost per task. The guide's last word: test the options on your own real work before committing.
For Fable specifically, Anthropic's Alex Albert suggests asking what's wrong with your idea before it builds anything, telling it up front what needs your approval, and handing it a whole project to run for hours rather than one step at a time. "This is where it shines," he wrote.
Why it matters: Most professionals pick an AI model by habit or by whichever is newest. Anthropic is telling customers that its flagship is overkill for most of what they do, which is unusual advice from a company selling it, and a useful reminder that the right setting depends on the task, not the leaderboard. The same logic applies to ChatGPT and Gemini: match the model to the job, and check whether a cheaper one does it well enough.
Sources: Anthropic (Claude developer guide) · Alex Albert on X
Companies Can Now Buy ChatGPT Enterprise Through AWS
ChatGPT Enterprise can now be purchased directly through AWS Marketplace, letting companies buy and bill it alongside their other AWS services and count the spending toward existing AWS commitments. OpenAI says customers get identical models and features to buying direct, with no lag on new releases. Specialized versions for healthcare and financial services are also available through the marketplace for eligible customers.
Why it matters: For large organizations already running procurement through AWS, this removes a budgeting and vendor-approval hurdle that can slow enterprise AI adoption, making it easier to expand ChatGPT access without a separate contract.
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Commerce Probe of the Alleged Spy's Firm Was Moved After Nvidia Reached Out, NYT Reports
China & Geopolitics Chips, Data Centers & Energy Nvidia
Commerce Department investigators were months into a case against Megaspeed, the Singapore company suspected of smuggling Nvidia's AI chips into China, when their work came to a sudden halt late last year, The New York Times reports. Megaspeed is led by Alice Huang, the executive U.S. officials believe is a Chinese spy and who had a relationship with Nvidia CEO Jensen Huang.
The case was abruptly transferred to the Justice Department after the agents got an email from Nvidia's general counsel, Tim Teter, seeking information about it. Teter wrote that Jeff Kessler, a top Commerce official nominated by President Trump and the agents' superior, had given Nvidia permission to reach out. The agents were alarmed, and some viewed Kessler's contact with Nvidia as violating government procedure and potentially tainting the case, according to officials who spoke to the Times. The case was moved to avoid compromising it, and the Justice Department is now conducting a criminal inquiry into Megaspeed, the Times reports. It's not clear what prompted Teter's outreach.
A Commerce spokesman said the department acted "strictly by the book." The Justice Department declined to comment. Nvidia spokesman John Rizzo said the company consults with officials "to answer their questions, provide information we have gathered from the field, request their guidance and instructions and ask whether they have any follow-ups."
The fallout from Huang's admission kept spreading. Nvidia's stock fell 4.27% on Monday, wiping about $250 billion off its value, the BBC reported. President Trump said the FBI is looking into the affair, without saying when the inquiry began; the FBI declined to comment. Trump said he called Huang on Monday: "I asked him that question, 'Did you have any idea she was a spy?' He said 'I had no idea she was a spy' and I really believe him." Also on Monday, the Rosen Law Firm filed a proposed shareholder class action in federal court in California, claiming Huang's failure to disclose the affair artificially inflated Nvidia's stock price, POLITICO reported. China called the allegations groundless.
Senator Elizabeth Warren, the top Democrat on the Banking Committee, posted that "Congress is back today. We have legislation ready to go to stop advanced AI chips from fueling China's military." On CBS's Face the Nation on Sunday, she said there is "bipartisan support in the Banking Committee" to limit chipmakers' exports to China. House Republicans' deadline for Huang to accept a Nov. 18 closed-door briefing is Wednesday; Nvidia hasn't said whether he will attend.
Why it matters: Until now the questions were about Huang's judgment. This report raises questions about the government's: whether a top Commerce official opened a channel for Nvidia into a smuggling investigation of one of its biggest customers, a company that bought $4.8 billion of its chips. That lands on Commerce, which decides who can buy Nvidia's chips, just as Democrats prepare to take over Congress with chip-export bills ready and subpoena power coming. For Nvidia, the risk is no longer one executive's private life. It's whether Washington keeps trusting the company it relies on to police its own sales.
Sources: The New York Times · Sen. Elizabeth Warren on X · CBS News (Face the Nation) · BBC · POLITICO
The Government's AI Safety Testers Are Leaderless, Quiet and Now Reporting to the White House
Regulation & Policy AI Safety & Risk
The small federal office that assesses the safety of new AI models has no permanent leader, has stopped publishing its research and has been drawn closer to the White House, the Associated Press reports. Officials at the Commerce Department division, now called the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), were recently asked to report informally to a White House office, two former government officials told the AP.
The center's director was hired and fired within days in April; his replacement resigned a few months later. It gets $10 million a year, about one-eighth of its British counterpart's budget. In August, Congress started asking why it had stopped publishing its reports on how safely AI systems perform. The White House didn't answer the AP's questions about the center, pointing instead to Trump's new Super Intelligence Force and a risk report he has commissioned.
Critics worry the center's findings could be politicized or muted in deference to the tech industry. Others argue that sitting closer to Trump could give it more prominence and perhaps more money. "At the end of the day, all of these decisions are of immense public consequence and should be explained and rationalized in the public eye," said Amos Toh of the Brennan Center for Justice. "And that's not happening."
Why it matters: The center is the government's main in-house check on the frontier models labs keep releasing, including the ones covered above. If its findings stay private and its staff answer to the White House, the public, and a Democratic Congress, will have a harder time knowing what the government has found, and the labs' own safety testing becomes the only scorecard.
Sources: The Associated Press, via WAFB
What's in the Lab
New announcements from major AI labs
CodeRabbit Says OpenAI's Codex Let It Ship 20x More Code Without Hiring
CodeRabbit, whose AI code-review tool is used by over 20,000 organizations, says OpenAI's Codex let it ship roughly 20 times more software over six months without adding engineers. The company's CEO says it now runs many more AI "subagents" checking code simultaneously, up from just one or two previously, and that OpenAI's GPT-6 Astra caught an estimated 50% of issues in its internal benchmark, topping other models tested. The team also rethought where human approvals belong, auto-shipping low-risk changes.
Why it matters: If a company can multiply output 20x by leaning on AI agents rather than hiring, it's an early data point for how AI reshapes engineering team size and structure, not just coding speed.
Manufacturer's AI Tool Turns Hour-Long Drawing Searches Into Minutes
Japanese manufacturer Kuken Industries, which builds cooling towers and ventilation systems, partnered with AI startup Recursive to build a search tool, powered by OpenAI's API, that lets engineers query a 70-year archive of equipment drawings in plain language. Kuken says the tool cuts search time from nearly an hour to about 4.5 minutes. Paper-era searches took half a day. Kuken projects a 90% reduction in annual hours spent hunting for drawings. Junior engineers can now find documents without relying on veteran staff's institutional memory.
Why it matters: It's a template for any old-line manufacturer sitting on decades of hard-to-search records. A narrow AI tool unlocked Kuken's archive and freed veterans to mentor newer engineers.
What's in Academe
New papers on AI and its effects from researchers
How Hospitals Deploy Discharge AI Can Entrench or Reduce Bias
A new study tackles a subtle feedback loop in hospital AI: when readmission-risk algorithms influence discharge timing, the data used to retrain those algorithms gets skewed, entrenching bias over time. In the researchers' model, simple "discharge when risk falls below X" policies can stay persistently biased. An alternative approach—factoring in bed capacity and letting stay lengths vary more—corrected itself over repeated retraining cycles. In simulations built from real hospital data, it cut prediction errors 24-25% and shrank cost disparities between demographic groups by up to 63%, without adding explicit fairness rules.
Why it matters: As hospitals increasingly let algorithms influence discharge timing, this research shows that how a model is deployed—not just how it's trained—can determine whether it gets fairer or more biased with use.
Simulated Crisis Chats Show AI Rushing Advice and Skipping Follow-Up
A new study tested a memory-equipped AI assistant across more than 14,000 simulated conversations built from 22 users' donated chat histories, examining how they respond to users navigating suicidal ideation, self-harm, and anxiety over time. Researchers found the assistant often gave advice too soon after a crisis, especially when recovery had stalled. It also frequently skipped follow-up and next-step planning that clinicians expect. The study, developed with mental health professionals, argues that checking single responses for safety isn't enough; timing and omission matter just as much as content.
Why it matters: As people increasingly turn to AI chatbots for emotional support, this research suggests the real risk isn't just what the bot says in one moment, but whether it tracks a person's trajectory well enough to know when to hold back or follow up.
Hospital AI Search Tool Wins Fast Clinician Adoption in Real-World Study
A study tracked 11 months of real-world use of Scout, an AI tool built into electronic health records that lets clinicians search and synthesize patient data with citations, across a multi-hospital academic health system. Adoption was fast and broad. 2,635 users ran 80,465 tasks, and daily sign-ups grew 21-fold at launch. Of practitioners who completed required training, 77.9% went on to use it, a median 1.7 days later. Use spanned 20-plus specialties and tasks from chart summaries to drafting notes, with a small group of power users driving most of the volume.
Why it matters: The findings suggest hospitals can get doctors and nurses to actually adopt AI tools at scale—but only when training is tied directly to access and governance scales with the tool's reach.
AI Models Struggle to Judge Teens' True Feelings in Chats
Courts & Lawsuits Kids, Teens & Mental Health
A new study tested ten AI reasoning models on whether they could infer mental health states from real Instagram DMs donated by 80 adolescents with histories of depression or suicidal thoughts, checking predictions against participants' own self-reports. The models did better at inferring observable states than subjective experiences, such as whether someone felt supported by friends. Researchers reviewed 204 of the models' reasoning traces to understand where the inferences broke down.
Why it matters: As schools, platforms, and health apps explore AI tools to flag at-risk teens from their online activity, this research suggests such systems may catch observable signals better than they read how a teen actually feels.
What's On The Pod
Some new podcast episodes
How I AI — Claude Fable 5.2 review: fast, creative, and finally not annoying
How I AI — How this founder built a better Duolingo with Codex | William Imoh
AI in Business — Getting Pharma AI Agents to the Right Answer by Fixing the Data and Semantic Layers - with Richard Knoche of Flatiron Health