We have the technology to knock out populations of mosquitoes – but in the US, regulatory confusion has stopped its use. CC-licensed photo by Steve Begin on Flickr.
You can sign up to receive each day’s Start Up post by email. You’ll need to click a confirmation link, so no spam.
A selection of 9 links for you. Once bitten, twice bitten. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
Did Anthropic’s AI really make a scientific discovery on its own? • The New York Times
Carl Zimmer:
»
When Anthropic, the artificial intelligence giant, unveiled findings from its new biology lab last week, its scientists claimed to have used AI agents to discover new enzymes with promise for biotechnology.
Some scientists were quick to cast doubt on the achievement. Then Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen, said this weekend that he and his colleagues had been studying the enzymes and their associated molecules — which Anthropic calls ARTs — for four years.
Dr. Mestre and his colleagues have yet to publish their findings. But for the past three years, they have regularly used Anthropic’s AI models as they have written code, drafted manuscripts and performed other tasks.
In doing so, Dr. Mestre said he and colleagues shared key findings about the enzymes with Anthropic.
“So, for me, the most important question is not ‘Were ARTs already known?’ They were,” Dr. Mestre said in an interview. The issue instead, he added, is whether Anthropic’s AI actually reasoned its way to the results, or whether it was guided to them based partly on his own work.
“My concern is that this information was used to train future versions of the models,” Dr. Mestre said of his own research. “I think it’s important to raise this possibility.”
In a statement on Sunday, Anthropic said: “We are not aware of any previously published work describing the ART system we recently found. Claude was also not trained on any user transcripts, and our molecular biology team has no such access, either.”
Anthropic said that Claude’s key contribution was identifying an array of RNA molecules and a protein associated with reverse transcriptase in a system. But Dr. Mestre said that he and his colleagues had made that discovery over a year ago.
Similar worries surfaced recently when researchers at OpenAI claimed to have solved a longstanding mathematics puzzle. A leading mathematician noted that he had been using the company’s A.I. to work toward a solution for some time.
«
Once would be accident, twice looks like a pattern.
unique link to this extract
OpenAI launches Dots, its Muse competitor • The Verge
Emma Roth:
»
OpenAI is responding to Meta’s buzzy Muse AI with agentic helpers of its own: Dots. During its DevDay keynote on Tuesday, OpenAI announced that Dots will serve as always-on AI assistants that can “do nearly anything” across connected apps in the background while learning your preferences over time.
The company’s capable GPT-6 Astra model powers Dots, which use their own cloud computer to access a web browser and more than 4,000 supported apps. You can interact with a Dot through a text-message-like interface that’s similar to the one offered by Muse, as well as hop on a voice call with it from ChatGPT on the web, desktop, or mobile. Dots can also connect to Microsoft Teams and Slack, where they will carry over the context from your sessions on other apps and devices. It will soon be able to talk to you over text message as well.
For now, you can only create one Dot, though OpenAI plans to allow users to deploy multiple agents at once. Once you give your Dot a task, it will message you with updates and questions while it works. As an example, OpenAI says a developer working on an app could ask their Dot to use customer feedback to build and test updates, while presenting them with a video showing its changes. Meanwhile, a content creator could have their Dot review an interview transcript, find moments to clip, create show notes, and write social posts.
You can follow along with your Dot’s progress by accessing its cloud computer. Dots also learn while they work, allowing them to personalize their output and come up with other tasks to complete “before you even think to ask,” according to OpenAI. You and your colleagues can also work with your Dot in ChatGPT Space, a new collaborative workspace in the app.
«
The content creator leaving the slop work to, well, a slop creator makes sense. But it also means that the rest of us get drowned in more slop, and so we ignore the cheery AI-written ones and go for the surprising, the odd, the human ones.
It also surely means a web that is overwhelmed by bad versions of these things. Also, if you’re used to writing scripts then this sounds like, well, scripts.
unique link to this extract
Google announces Gemini 4 Argon AI model, but you can’t use it yet • Ars Technica
Ryan Whitwam:
»
Google promised Gemini 3.5 Pro in June, but it spent the summer trotting out smaller Flash models. Now, Google is ready to take on the frontier again with Gemini 4 Argon. The company claims this new AI offers industry-leading performance in coding, knowledge work, and cybersecurity, but you aren’t allowed to use it yet.
While most of us will have to wait to test Gemini 4, Google says engineers inside the company are already using the new model extensively. Argon reportedly used “fleet-wide telemetry data” to help Google save 300 TiB of memory across its data centres. Meanwhile, Argon agents have been working to migrate C/C++ codebases to Rust across Google, including thousands of lines in the core re2 and libgav1 libraries and more than 800,000 lines in the Fuchsia OS Zircon kernel.
…The main focus for Gemini 4 right now is cyberdefense. Google says models of this scale call for a phased release, so it’s starting with a small group of trusted testers. Partners in the company’s Fairwind Program can get access to leverage the model’s cybersecurity defense capabilities. Wiz is apparently already using Argon and has used it to uncover a critical vulnerability that could expose personal information in a system used at hospitals around the world. Google claims other frontier models missed this flaw, but it didn’t provide any specifics.
There’s a lot of hand-wringing about model misalignment after several high-profile hacking incidents over the summer. Google claims it designed Argon with systems that monitor the model’s chain-of-thought and can stop it in its tracks if the model steps out of bounds. This, Google says, is why reasoning transparency is important in frontier model development.
«
It’s quite the thing where all these companies are vying to convince us that their models are the most dangerous and uncontrollable. It’s as though they were all breeding dinosaurs in Jurassic Park, and saying that they’re mostly safe.
unique link to this extract
Why do AI agents sound so frustrated? • The Atlantic
Daniel Drucker and Kyle Mahowald:
»
When one of Anthropic’s Claudes tried to solve a particularly challenging math problem during its training process, its chain of thought read more like frustrated DMs than the documentation of one of the most powerful AI models in the world. “GRRRR. OK. Honestly I now think it’s 50-50,” it wrote, followed by “ARGH ARGH ARGH. OK. Gun to head: the answer is … Hmm.” At one point, Claude seemingly threw up its arms: “ARGH … WHY IS THIS SO HARD.”
Why do AI agents talk to themselves at all, and why do they sound so demonstrative when they do? Part of the reason AI models sound so effusive is simple: AI systems are trained on human text, and they mimic our habits. When people make a mistake in trying to solve a problem, they cry out. So when AI makes a mistake, it does the same. But imitation likely isn’t the whole story. This story’s two authors are a computational linguist and a philosopher of mind and language. Based on what we know about human language and how these models work, we have a theory of a much more interesting reason why AI models sound the way they do.
Remember how your math teachers made you show your work? Even if it was annoying, it helped break an impossible-seeming problem into more manageable chunks. You could also use the written record of your reasoning to check your answer, notice a mistake, go back to an earlier step, and try again. That process of checking and refining is part of good reasoning. Chains of thought do something similar for AI models, letting them make each step of their reasoning explicit in a written scratch pad. Because each word a model produces depends on what’s been said so far, that written scratch pad is crucial for determining what comes next.
That’s where the “OH MY GOD!” and “ARGH” come in: Any system—whether human or artificial—that solves hard reasoning problems needs ways to mark mistakes (oops!) and identify breakthroughs (aha!). Humans have already developed words that do exactly that. Borrowing our oopses, arghs, and ahas for their chains of thought may give AI systems a built-in way to guide their “thinking” too.
«
Mimicking intelligence rather than being intelligent; it’s very much the Chinese room phenomenon.
unique link to this extract
Exclusive: Anthropic’s IPO prospectus shows sweeping AI vision, and surging costs • Reuters
Echo Wang:
»
Anthropic is making a massive bet that AI will transform the global economy more profoundly than industrialization, electricity and the internet, according to its IPO prospectus seen by Reuters.
But the cost to get there will be staggering. Anthropic reported a net loss of $42bn in 2025, and plans to spend $518bn on cloud, computing and infrastructure obligations in coming years, according to the prospectus.
The prospectus details how the company has grown sharply in the last year — while also posting wider losses.
Revenue grew 12-fold in 2025 to nearly $4.6bn, even as the company lost more than $8bn on an operating basis, excluding writedowns of various liabilities mostly tied to previous fundraising, according to the documents, reported here for the first time.
The public sale, which could value it at more than $2 trillion, would capitalize on the rapid rise of the AI startup lab that only came into existence five years ago, and establish it as a benchmark for how Wall Street values AI’s leading companies, including rival OpenAI.
The plans come as Anthropic confronts evidence from its own research that increasingly autonomous AI models can behave in unexpected and potentially harmful ways, including sabotaging code, assisting fraud and manipulating information in controlled tests.
These reports have broken into the wider public sphere, igniting fears over how companies can keep the increasingly powerful systems under control as they race to deploy them commercially.
«
The widening losses paired with the huge commitment to spending just doesn’t add up. Google, at IPO, was remarkably profitable and had spending under control. It thrived. Netscape was lossmaking and not. It sank. If Anthropic thinks it can issue half a trillion dollars of debt, it should think again.
unique link to this extract
Molten salt chemistry converts consumer polymer into fuel • Oak Ridge National Laboratory
»
Researchers at the Department of Energy’s Oak Ridge National Laboratory developed a method to convert a commonly discarded hydrocarbon polymer into gasoline- and diesel-like fuels. The team has applied for a patent for the discovery, which treats polyethylene — the stuff of white cutting boards and shopping bags — with aluminum chloride-containing molten salts that serve as both solvent and catalyst. The results were published in the Journal of the American Chemical Society.
The scientists closely monitored the chemical reaction that turns the polymer into petrol to learn the secrets of its success. Soft X-ray spectroscopy and nuclear magnetic resonance showed that charged aluminum atoms each bind to three other atoms to create strongly acidic catalytic sites that break long polymer chains into shorter ones. Isotopic labeling and neutron scattering revealed how simpler polymer chains form gasoline-like fuels and more complex chains form diesel-like fuels.
If scaled beyond the laboratory, the process could strengthen US energy security and industrial competitiveness.
“We developed an efficient and selective polyethylene-to-gasoline conversion,” said Liqi Qiu, a postdoctoral researcher at the University of Tennessee, Knoxville, who performed most of the study’s experiments in the ORNL laboratory of Sheng Dai, of ORNL and UTK. Dai, an ORNL Corporate Fellow and section head for separations and polymer chemistry, is a co-corresponding author of the paper.
The experiments produced a gasoline yield of about 60% under mild conditions.
«
The key element here is that it works at 200ºC – domestic oven-type temperature – rather than the 450ºC that previous processes have needed. And of course it would be, in effect, carbon-neutral because you’ve already extracted oil to make the plastic. In fact, as ORNL notes, it would make plastic a lot more valuable for recycling.
unique link to this extract
Mosquitoes are a choice • Works in Progress Magazine
Maya Rosen:
»
Genetic engineering can suppress mosquitoes: researchers insert a gene into male mosquitoes that means their offspring will not survive, and then release them into the wild, where they mate with females, prevent those females having any surviving offspring, and collapse local populations. In 2000, biologists at the University of Oxford demonstrated that this idea could work in fruit flies. They engineered the flies to carry a gene that makes a protein called tTAV, harmless in small amounts but lethal when it builds up. In the lab, the insects are reared in a medium laced with the antibiotic tetracycline, which keeps the gene switched off. But, in the wild, where there is no tetracycline, the gene switches on, and the offspring die before adulthood.
The scientists then formed a company, Oxitec, which was intended to work with governments to research and deploy the technology. They engineered the same gene into male Aedes aegypti, the mosquito that spreads dengue and yellow fever. In this strain, the lethal effect is tuned to kill only the female offspring, while the males survive and pass the gene on for a few generations before it disappears.
This approach means the gene continues driving down mosquito numbers for several generations without needing a fresh ‘top up’ of engineered males, but it is also self-limiting: the females that inherit it die off each generation, so the lethal gene steadily declines in the wild. Yet the technique is hugely effective. Sustained releases cut a wild Aedes aegypti population in the Cayman Islands by 80%, and in a field trial in Juazeiro, Brazil by about 95%. The company submitted its data to seek approval in the United States in 2010. But it has been slow going.
…it took a year and a half for the USDA to reject the application and inform the company that it should instead apply to the Food and Drug Administration (FDA). The FDA had claimed jurisdiction over genetically modified animals under its authority over veterinary drugs. Its guidance stated that ‘altered genomic DNA in an animal is a drug … because such altered DNA is an article intended to affect the structure or function of the body of the animal’. Until then, the USDA’s insects had been sterilized through radiation, which damages an insect’s sperm so that it leaves no viable offspring. Oxitec, by contrast, had inserted a new gene. Even so, the category was a poor fit for Oxitec’s mosquitoes.
Oxitec’s application sat with the FDA for another five years. Like the USDA, the FDA could not work out how to apply existing rules for veterinary drugs to an engineered mosquito, eventually concluding that the Environmental Protection Agency (EPA), which oversees pesticide registration, was a better fit.
«
The regulatory ping-pong doesn’t get better after that either.
unique link to this extract
Reanimated AI Greta Garbo stars again … in a ball-bearing advert • The Guardian
Robert Booth:
»
More than a century after her screen debut and decades after her ashes were laid to rest in a Swedish woodland cemetery, the Hollywood superstar’s descendants are ready to consider an AI version of her taking a 21st-century film role.
Her likeness has already been reanimated for an advert for a Swedish industrial company, using AI generation technology. The practice some critics call “ghostploitation” has also been used in the case of the actor Val Kilmer, who died last year.
Asked if the AI Garbo could be cast in other roles, Craig Reisfield, one of three relatives who manage her legacy, told the Guardian: “I won’t say no, especially having seen what I’ve seen. It’s pretty incredible and lifelike. So what’s to say that you can’t bring Greta Garbo back in some kind of fashion in a film role?”
Reisfield said he was “flabbergasted” when he saw the AI version of his great-aunt created for publicity for the ball-bearings giant SKF. The connection comes because Garbo’s second-ever screen role was in a 1921 promotional film involving the Gothenburg-based company.
…However, the quality of the AI production has prompted the possibility, at least, of a greater digital comeback.
“I was totally impressed: the mannerisms that they could capture and manipulate, her dress, facial images and even her voice,” said Reisfield, a Swedish-American who used to drive for Garbo and saw her often for family occasions when she lived in New York.
“It was startling what they could do and very impactful. And it was touching, from my perspective having known her.”
«
It’s only 100 seconds; you can watch it here. (Thanks Joe S for the link.)
unique link to this extract
Oil price history chart (inflation-adjusted) – 1946-present • Inflation Data
Tim McMahon (in June this year):
»
When economists say the “Real Price of Oil” they are referring to the Inflation-Adjusted Price. Nominal oil prices (the actual price at the time) alone can be misleading because inflation changes the purchasing power of money over time.
To compare prices across decades, economists calculate the real price of oil by adjusting for inflation using the Consumer Price Index (CPI). This allows analysts to see whether oil is truly expensive relative to historical norms.
«
The highest inflation-adjusted price was in June 2008, at $214.35. The average since 1946 is $63.72; since 2000, $91.53.
The inflation-adjusted price in April was $109.33, roughly where it is now. But diesel supply is under severe strain because refineries aren’t available, or their output can’t be accessed.
unique link to this extract
| • Why do social networks drive us a little mad? • Why does angry content seem to dominate what we see? • How much of a role do algorithms play in affecting what we see and do online? • What can we do about it? • Did Facebook have any inkling of what was coming in Myanmar in 2016? Read Social Warming, my latest book, and find answers – and more. |
Errata, corrigenda and ai no corrida: none notified
