Eight X signals: Claude finds a new enzyme system, Google ships a voice studio, and the web braces for agent swarms

In the 24 hours ending September 24, 2026, seven of the fourteen AI and tech accounts this channel tracks produced something worth carrying, and every item below starts from one of them. Research carries the day: Anthropic's new biology lab reported the first discovery its Claude agents made on their own, OpenAI and Epoch AI each opened a benchmark to outside use, and Google put voice design into the hands of anyone with an API key. Retweets without comment, replies, jokes, and posts about something other than AI or technology are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, has left X, so its posts are unreadable and absent from this edition.

Google's new speech models design a voice from a prompt, then direct it line by line Google blog post · announcement on X: 460 likes, 69,886 views at capture · September 23

Google released two text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, which replace a fixed preset list with a voice you describe 1. The Flash model builds a voice from a natural-language prompt across more than 100 languages and dialects, and it draws on a library of 2,000 production-ready voices where the previous generation shipped 30 [cite:1]. Voice replication rebuilds a speaker from a 30-second sample, and it requires a verbal consent recording from the voice's owner that matches the reference before the voice can be created [cite:1].

Both models take stage directions written into the script, so pacing, emotion and dialect shifts can be controlled line by line, and both can stage a two-speaker conversation from a single script with laughter and other backchannel sounds written in [cite:1]. Google says Flash TTS takes first place on Hume AI's Voice Design Benchmark at 71.4 and leads that benchmark's accent modelling at 60.8, and that the two models hold first and second place on Hume's overall quality index [cite:1]. Every clip carries a SynthID watermark, Google's inaudible marker for AI-generated audio [cite:1].

Simon Willison measured what the cheap end costs. He built a small bring-your-own-key playground for the API and generated 1 minute 18 seconds of audio in about 20 seconds for 2.74 cents, using Flash rather than the cheaper Flash-Lite 2. One script can carry a whole conversation between characters with different voices and their own style instructions, which is what he used it for: two pelicans debating a move to Pacifica Pier [cite:2].

ChatGPT Voice can now run your email, calendar and Slack by talking X post · 10,913 likes, 2,096,678 views at capture · September 23

OpenAI turned ChatGPT Voice into something that acts inside other applications. The assistant can now use plugins for email, calendar and Slack, run on GPT-6 Astra, Sol and Luna, and work inside ChatGPT Work on the web and on mobile, so a spoken request can produce a document, a deck, a site or a spreadsheet 3. OpenAI said the change rolled out globally in the app that day [cite:3].

Greg Brockman, OpenAI's president, called it a "mega upgrade for GPT Voice, which can now use tools and is available in Work" 4.

Simon Willison pointed it at something of his own within hours, holding a voice conversation with the Datasette backup of his blog from the ChatGPT iPhone app 5. Willison builds Datasette, a tool for publishing and querying data, and datasette-mcp is the connector that puts a Datasette database in front of a model; that pairing is what let a spoken question reach his own archive [cite:5].

Claude agents found a new enzyme system in bacterial viruses X post · 36,742 likes, 21,121,207 views at capture · September 23

Anthropic formed a life-sciences research group in the spring of 2026 to test whether general models can systematise biological discovery, and built its own molecular biology laboratory in the Bay Area to run the work. The lab handles only the two lowest biosafety levels, and human scientists do every bench experiment 6. Its first reported result is an enzyme system that Claude found on its own, inside bacteriophages, the viruses that infect bacteria [cite:6].

The system sits next to a long array of evenly spaced DNA repeats, a layout that resembles the array at the heart of CRISPR, and it is built on a reverse transcriptase, an enzyme that copies RNA back into DNA [cite:6]. Anthropic named it ART, short for array-associated reverse transcriptases. Claude gathered more than 200,000 reverse transcriptases, narrowed them to 3,500 candidate systems and then to the 20 most compelling, across 21 hours of searching by roughly 950 agents that spent 210 million tokens [cite:6]. What the system does remains an open question; the first experiments show that its repeat array is expressed as a set of distinct short RNAs [cite:6].

Feng Zhang, the MIT and Broad Institute professor who was one of the pioneers of CRISPR genome editing, reviewed the pre-print and called the work "an exciting example of how AI agents can contribute to biological discovery" [cite:6]. Anthropic published the pre-print alongside the announcement and invited outside scientists to send in research questions for the group to take on [cite:6].

OpenAI opens a mental-health benchmark written by 80 clinicians X post · 3,808 likes, 511,871 views at capture · September 23

OpenAI published MentalHealthBench, an open benchmark for how models handle mental-health conversations, built with more than 80 licensed psychologists and psychiatrists across 22 countries who speak 19 languages and cover nearly 20 subspecialties 7. Its conversations are synthetic, built with privacy-preserving techniques, and they spread across three levels of need: everyday emotional conversations, serious distress short of an emergency, and emergencies [cite:7]. The personas include adults, teenagers aged 13 to 17, caregivers and clinicians; OpenAI says more than one billion people use ChatGPT each week [cite:7].

The scoring is where this benchmark departs from earlier work, which concentrated on emergencies. Each expert read a conversation and wrote rubric criteria for the model's reply to the final user turn, weighted from minus 10 to plus 10, with negative points for harmful behaviour and larger values for greater clinical importance [cite:7]. Every conversation was reviewed by at least three experts, and only criteria that two agreed on and a third did not contradict survived into the benchmark, which an automated grader, GPT-5.6 Sol, then uses to score model responses [cite:7].

Arthur Evans, chief executive of the American Psychological Association, framed why the range matters: "Mental health exists on a continuum, from flourishing to everyday stress to acute crisis" [cite:7]. OpenAI released the benchmark and its method openly so other researchers can run their own evaluations against it [cite:7].

The cost of a given level of AI performance is falling about 13 times a year Research publication and X posts · benchmark post 1,002 likes, 95,440 views; cost post 240 likes, 16,140 views at capture · September 23

Epoch AI, a research group that tracks the trajectory of AI, measured how fast the cost of a given level of performance falls across five benchmarks covering mathematics, science and games of skill. Its answer is about 47% per quarter, or 13 times a year, since 2023, a faster decline than electricity, compute, batteries or DNA sequencing ever managed 8. On September 23 the group added a caveat to that headline: "the price of thought" is a looser notion than the price of a kilowatt-hour, which makes comparisons between them harder than the chart suggests [cite:8].

The same group published the capability side of the ledger that afternoon. Its Furniture Assembly Benchmark hands a model the assembly manual and a photograph of a half-built flat-pack piece of furniture and asks it to spot the mistake, and the top score climbed from 28% to 80% in ten months 9. Ethan Mollick read it as "a useful measure of how much better multimodal AI has gotten in a short time" — multimodal meaning a model that reads images as well as text 10.

Mollick drew a warning out of the cost curve. Costs are collapsing while measured ability rises, he wrote, which is why "optimizing for cost for a particular solution right now may end up being shortsighted" 11.

Claude is doing the daily arithmetic of the Ebola outbreak in eastern Congo X post · 670 likes, 90,232 views at capture · September 23

Health teams in the eastern Democratic Republic of the Congo have been fighting Bundibugyo ebolavirus since May, a rare strain for which no vaccine is approved. The one licensed Ebola vaccine, Ervebo, was made for the Zaire strain. Almost half of the nearly 8,000 confirmed cases have died 12. A partnership convened by the Coalition for Epidemic Preparedness Innovations, with the WHO Regional Office for Africa and the DRC's National Institute of Biomedical Research, has been using Claude across the response [cite:12].

The clearest saving is the situation report, the daily document responders use to decide where Ebola is heading and where to send help. A WHO AFRO data team built a Claude skill that pulls case and laboratory numbers out of each health zone's PowerPoint deck, checks them against the previous day's report, flags any change in trend, and summarises the district reports; work that took a full day now takes under an hour [cite:12]. The same team runs several disease models at once where time constraints previously allowed one, and uses the resulting forecasts to help logistics staff decide where treatment centres should be built [cite:12].

At the INRB laboratory, which has sequenced essentially all of the outbreak's genomes, Claude Science, an AI workbench for scientists, assembles a virus genome from millions of short fragments and builds its family tree from a plain-language prompt, which lowers the bioinformatics barrier during an outbreak [cite:12]. Polina Brangel of CEPI described where the boundary sits: choosing which cohorts of serological samples to analyse for Ervebo's potential cross-reactivity against BDBV stays a scientific judgement made by experts, while Claude organises complex, multi-factor data so those experts can compare a wide range of vaccine proposals in a much shorter time [cite:12].

Superforecasters gave AI a 1.7% chance of cracking a Millennium Problem by this month X post · 693 likes, 33,926 views at capture · September 24

Ethan Mollick looked back at what forecasters had said in September 2025 and found the best superforecasters had put the chance of AI resolving a Millennium Problem by September 2026 at 1.7%, with the more optimistic industry expert at 4.6% 13. Those same forecasts greatly underestimated the revenue AI labs would earn, he noted [cite:13].

The Millennium Prize Problems are the seven one-million-dollar questions the Clay Mathematics Institute set in 2000, and Navier–Stokes, the behaviour of fluid-flow equations, is one of them. OpenAI announced on September 8 that its model had solved it, and mathematicians who have read the 166-page manuscript say they believe the proof is technically correct 14. What they have taken from it so far is close to nothing: "The paper is not written for humans," said Javier Gómez-Serrano of Brown University, who believes a serious rewrite could make it useful to the field [cite:14].

Mollick credited the forecast numbers to a thread from the Forecasting Research Institute. Its earlier round, run in the summer of 2022, had most forecasters underestimating AI progress on three tests — MATH, for competition mathematics, MMLU, for broad knowledge, and QuALITY, for long-passage reading — by mid-2024, and it found that AI reached International Mathematical Olympiad gold standard in July 2025, five years earlier than the median expert prediction and ten years earlier than the median superforecaster's 15. Those first forecasts were collected before ChatGPT shipped, and the Institute's September 23 thread asks whether forecasters are making the same mistake about the years since [cite:15].

Paul Graham sees several startups in the fight over agent traffic X post · 231 likes, 22,578 views at capture · September 24

Nikita Bier, X's head of product until recently, wrote that bot detection and human verification will be one of the most urgent demands on businesses in the coming years, that agent swarms will "suffocate every website and form", and that small companies and government websites are the most exposed 16. When X went looking for a vendor that combined the current techniques, he said, it found none and built the system in house [cite:16]. The post drew 11,890 likes and 1.1 million views from other readers weighing the same problem [cite:16].

Paul Graham read it as a supply gap with room for more than one company in it. "There will probably be one variant meant to filter out agents, but another meant to welcome them in an orderly way," he wrote, adding that the problem is large enough to build several businesses on 17.

Another of Graham's posts that morning made the pricing version of the argument. A lab whose models alone can solve really hard problems keeps its pricing power, he wrote, because "people will pay more per token to solve really hard problems" 18. The claim lands two days after Anthropic and OpenAI each released a cheaper model on the same afternoon 19.