Ironic indeed
Gags on Musk, brain-computer interfaces, Talkie 1930, verified humans, and more
Dispatches from Mitch
Musk vs. Altman, Day 3
Despite the fact that OpenAI’s creation was largely motivated by concerns about the extinction threat of artificial superintelligence, Judge Yvonne Gonzalez Rogers has declared the topic off limits. Major news outlets noticed.
The Guardian quoted the judge as saying, “We are not going to talk much about extinction in this case. They got it, that’s enough.” This was in response to Musk saying, “The worst-case situation is where it is a Terminator situation, where AI will kill us all.”
The New York Times added a remark from her that, “I suspect that there are a number of people who do not want to put the future of humanity in Mr. Musk’s hands. But we’re not going to get into that.”
Reuters relayed the judge’s rationale for barring Musk’s lawyer from introducing expert testimony on the topic: “[T]his is not a trial on the safety risks of artificial intelligence.” She noted that it was “ironic that your client, despite these risks, is creating a company that’s in the exact same space,” referring to xAI.
This may upset the prosecution’s plans; AI risk expert Stuart Russell was expected to be one of Musk’s team’s next witnesses.
I’m not sure anything else in the trial matters as much as these discussions (or their absence), but sure, there seemed to be some grinding progress toward an eventual verdict. In words the judge has barred Musk from continuing to repeat, Musk accuses Altman and company of having “stolen a charity” by converting the valuable parts of the non-profit Musk helped start into a for-profit enterprise, against OpenAI’s charter, and on exploitative terms.
Under cross-examination, Musk admitted he “didn’t read the fine print, just the headline” of the 2017 term sheet that started the first phase of conversion. OpenAI claims Musk is just bitter about the company’s success, as someone who had tried and failed to control it before starting a competitor.
Hit the road, jack in?
In a world where the race to artificial superintelligence were halted by international agreement, I would expect many more of our headlines to be like yesterday’s Reuters story about Charles Lieber, a Harvard chemist who fled to China to run a state-funded brain-computer interface lab.
In 2021, Lieber was convicted of lying to U.S. authorities about payments from China received for helping them recruit overseas talent. He spent two days in prison and six months under house arrest. With court approval, he then started taking “employment networking” trips to China.
The lab he now runs, i-BRAIN, has its own chip-making hardware and access to a 2,000-cage primate facility on the same campus. Yes, this is about wiring biological brains into computers. Near-term applications include restoration of mobility for those with paralysis. Longer-term, human augmentation is the goal.
In March, China named brain-computer interfaces tech a national priority. The People’s Liberation Army has explored brain interfaces for “super soldiers”, and so has U.S. military tech incubator DARPA, which is especially interested in drone and cyber applications. The Pentagon had actually put over $8 million into Lieber’s Harvard research before the conviction.
If you look at Lieber’s case and wonder if his conviction was counterproductive to U.S. interests, you’re not alone. The enforcement program that had snared him was discontinued for failing to have the desired effects.
I say these stories would be more common in a world where the AI race wasn’t so all-consuming because as sci-fi as it sounds, brain-computer interfaces show a lot of long-term promise. I could imagine the field progressing a lot faster if it saw investment remotely comparable to that going into AI research. There’s plenty of dystopian potential with either tech, but if the first minds capable of outmaneuvering humanity were still essentially human, I’d like our odds a lot more.
Two scoops of resentment
We should not be surprised that journalists would take an interest in stories about AI tools augmenting or replacing journalists. The New York Times reported today on a byline strike at McClatchy, a 30-paper chain using an internal tool called the Content Scaling Agent.
The Agent summarizes a source article and generates different versions for different audiences. (The irony of my own reporting on this is not lost on me.)
What’s a byline strike? Reporters at nine papers are refusing to have their names attached to the Content Scaling Agent’s outputs, even in a caveated form that looks like, “Produced using A.I., based on original work by [writer].”
The company claims they need the human bylines to show “authority” to search engines, which feels extra sad to me: Humans who have had zombie articles made out of their material are being asked to lend their identities to machine outputs in order to impress other machines.
AI is eating India’s film industry
The Hollywood Reporter makes a strong case for India being the place to watch if you want to understand what happens when a film industry has few cultural or institutional barriers to AI integration.
Last August, the studio that owns a cult favorite romantic tragedy, 2013’s Raanjhanaa, used AI to replace the ending with a happy one where the love interest lives. This was over the objections of its star and director. There was some public outcry, but many went to see the new version, and some openly declared they preferred the happy ending.
The piece goes on to describe studios producing lavish visual effects on shoestring budgets; the results might not be on-par with the best of Hollywood, but they’re not too far off, and audiences seem to like them.
Actors’ unions are essentially non-existent in India, where audiences are fragmented into many languages, each with their own microcelebrity voice dubbers.
Most of the 20,000 voice artists in that dubbing industry look to be out of work soon, given estimates that 70-80% of India’s TV and video commercial brand voices are already AI-generated.
A pleasant journey to you, sir
The consistently great Hard Fork podcast has a fresh installment with segments on OpenAI’s disentanglement from Microsoft, the Musk vs. Altman trial, AI in healthcare, and a “vintage” AI model called Talkie 1930.
All are worth a listen, but it’s that last one I want to zoom in on here. We haven’t covered Talkie yet, and it’s been a darling of AI Twitter this week. The Hard Fork hosts chatted with one of its creators, David Duvenaud.
Talkie 1930 is a large language model trained only on writing and other data published before 1931. Why 1931? Because essentially everything older than that is in the public domain.
Duvenaud explains that the charmingly literary-sounding chatbot is intended to help study the intrinsic abilities of AI models to make discoveries and forecasts from limited data. This is a lot easier when researchers know “future” discoveries and events that the model doesn’t.
Talkie 1930 isn’t quite ready for that, though. Duvenaud points to some current issues: For one, the model itself (not just the dataset) is too small to be intelligent enough for coherent back-and-forth chat. For another, “There’s definitely contamination” they need to root out; many archival files in the data set are mislabeled, or are from newer editions that include anachronistic notes, etc., so Talkie knows things it shouldn’t. And Duvenaud’s team hasn’t even tried to suppress the model’s hallucinations.
The model can also be pretty racist and sexist, reflecting the values on display in its training corpus. Trying to suppress this would have destroyed some of the value of the experiment, so they’ve left it as-is for study, and added a modern overseer AI that flags potentially upsetting content for any users who’d rather pass on it.
When host Kevin Roose asked, “What’s a good goodbye to a podcast guest? Don’t worry about what a podcast is,” Talkie replied, “A pleasant journey to you, sir.”
Co-host Casey Newton asked the more important question, though: “If you put the model in a robot, would that be a walkie-talkie?”
You can chat with Talkie 1930 at talkie-lm.com.
Dispatches from Stefan
Six months to replace Claude
Today, we observe the Pentagon’s relationship with AI being rewired in real time. Many major outlets are covering the same announcement from slightly different angles: the Pentagon has now completed classified-work agreements with seven AI companies — OpenAI, Google, SpaceX, Microsoft, Amazon (AWS), Nvidia, and a startup called Reflection AI (backed by Nvidia). All seven signed onto the “any lawful use” standard that Anthropic had refused. The whole arrangement is described as a deliberate transition away from Anthropic’s Claude, which currently powers Maven Smart System, an AI tool involved in Iran operations and which the Pentagon has given itself six months to replace.
However, the most concerning piece of information comes from Bloomberg’s Katrina Manson: Nvidia’s specific agreement licenses its models for use in autonomous weapons systems development, and the company further agreed not to impose any usage policies “beyond what is required by US law and constitutional authority,” with no clearly stipulated monitoring or evaluation mechanism attached. That happened the same week Defense Secretary Pete Hegseth told Congress: “We follow the law and humans make decisions. AI is not making lethal decisions.”
Manson is direct in describing what the Pentagon did to Anthropic. The department, she writes, “refused to heed Anthropic’s stated red lines” — mass domestic surveillance and fully autonomous weapons — and “sought to eject the company from all defense supply lines.” Hegseth, in the same Congressional testimony, called Anthropic CEO Dario Amodei an “ideological lunatic.”
A couple of oddities are worth flagging. Reflection AI — the surprise name on the list — is a startup founded by ex-DeepMind researchers, backed by Nvidia, in talks at a $25 billion valuation — but has yet to release a public model. Meaning, the Pentagon is now signing classified contracts with AI companies that don’t have shipping products.
Meanwhile, the Anthropic story keeps getting stranger: Pentagon staffers told Reuters they consider Claude “superior to alternatives” and are reluctant to phase it out, and the New York Times reports officials are quietly pushing for a compromise that would let other parts of the government use Mythos, Anthropic’s newest model. The lesson the industry just learned is a bad one: red lines cost you the contract, but capability gets you back in the room. That’s not how you get safer AI deployment. That’s how you get companies that learn to keep their objections internal and let their models do the talking.
Verified human, maybe AI
The BBC reported today on Spotify’s new “Verified by Spotify” badge, awarded to artists who show “signals of a real artist behind the profile”: touring, merchandise, interviews, and the like. The badge certifies the artist is human. It says nothing about whether the music was made with AI. Spotify seems like it might be hoping people don’t notice the distinction.
Ed Newton-Rex, a creator-rights campaigner and former AI executive, suggests a simpler fix: Just label AI-generated music as AI-generated. The cautionary tale is a band called The Velvet Sundown: a “band” that had a verified Spotify page and 850,000 monthly listeners in 2025 before users realized they’d never given an interview or played a show. The new badge would not have caught that. A label saying “this music was made by AI” would have.
Triage test
The Guardian reported yesterday on a Harvard study in Science finding that OpenAI’s o1 reasoning model (now considered very out of date) outperformed ER doctors at triage diagnosis (67% vs. 50–55%) and crushed them on long-term treatment planning (89% vs. 34%). In one case, a patient with a blood clot in the lungs seemed to be failing on anticoagulants — the AI flagged the patient’s lupus history and correctly attributed the inflammation to that. The doctors did not.
But University of Sheffield’s Dr Wei Xing warns in the same piece that doctors may “unconsciously defer to the AI’s answer rather than thinking independently.”
Nearly 1 in 5 U.S. physicians are already using AI for diagnosis; in the UK, 16% use it daily and another 15% weekly. So the doctor-plus-AI setup The Guardian pitches isn’t a future model.
Will doctors keep doing the harder cognitive work themselves once an AI is reliably available to do it for them? Diagnostic skill, like any skill, atrophies without use. A generation of doctors trained to defer to AI by default would be cheaper to employ and faster at common cases, but worse at handling the rare conditions, the unusual presentations, the patients whose symptoms don’t fit any pattern. This is where you want a doctor who’s stayed in the habit of thinking, not one who’s spent a decade rubber-stamping AI output.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual analysts and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.