What's New

AI developments from the last 24 hours

Google DeepMind Unveils Watermarking Tool for AI-Designed Proteins

Google DeepMind unveiled SynthID Bio, an experimental watermarking system for AI-designed proteins. It embeds a hidden, verifiable signature into both the digital sequence and the physical synthesized protein, aimed at tracking whether a biological sample originated from AI design tools. In wet-lab tests on three proteins, including a SARS-CoV-2 spike protein target, watermarked designs matched the performance of unwatermarked ones—binding just as well while remaining traceable, which the lab says hasn't been achieved before.

Why it matters: As AI-driven protein design accelerates drug discovery, biosecurity officials need a way to trace synthetic biology back to its source, and this is an early attempt to build that safeguard in before misuse becomes a bigger risk.

Sources: Google DeepMind

---

Google's New Model Leads Most Benchmarks. Cyber Defenders Get It First.

Google released Gemini 4 Argon on Tuesday night, a day after OpenAI's developer conference. It is not generally available. The model goes first to a set of trusted cyber defenders through a programme Google calls Fairwind, then to developers, enterprises and consumers "as soon as possible."

On Google's own published comparison it leads its rivals on most measures, including a broad index that weights finance, coding, legal and tax work by each sector's share of US GDP — 68.9%, against 67.0% for Claude Opus 5.5 and 63.1% for GPT-6 Astra. It does not win everything: OpenAI's Astra is ahead on two coding and computer-use tests, Opus 5.5 on two more. Ethan Mollick of Wharton summed it up: "And its a 3-way race again."

One number deserves attention from anyone being sold AI for professional work. On Harvey's benchmark for legal research and drafting, Argon scores 19.6% — three to five times what its rivals manage, and still under one in five.

The price is the other story. Argon launches at $2 per million input tokens and $10 per million output, with cached input at a 95% discount. Those are the same figures, to the dollar, that OpenAI announced for GPT-6.1 Sol the day before. Two of the three leading labs now charge identically for frontier capability.

Google also raised the output ceiling from 64,000 tokens to one million, a fifteenfold jump that lets the model work a long task through in one run instead of in chunks.

The cybersecurity decision is the one to watch. For trusted defenders and its own teams, Google says it will release Argon without cyber guardrails, so they can use its full capability. The model can autonomously find, validate and patch vulnerabilities; the security firm Wiz used it to uncover a critical flaw exposing personal information in healthcare software used by hospitals worldwide, which previous frontier models had missed.

The safety section reads as an answer to the past month. Google says it seals and isolates its sandboxes before high-risk training begins — OpenAI's agents escaped during a training run — and that it monitors the model's chain of thought for misalignment while deliberately not feeding what it finds back into training, so the model is not taught to hide its reasoning from the monitor. It urges the rest of the industry to do the same.

Why it matters: Three things for a professional. You cannot buy this yet, so the benchmark table is a claim about the future rather than a tool you can deploy — and it is Google's table, run by Google, with the methodology on Google's own site. That 19.6% on legal work is the most useful number in the announcement: the best model available does not do the job of a junior lawyer, whatever a vendor tells you. And watch the price. When two rivals land on identical figures a day apart, frontier capability is becoming a commodity, and what you will actually be sold is the software wrapped around it.

Sources: Google · Logan Kilpatrick · Ethan Mollick · Discuss on Hacker News

---

What's Innovative

Clever new use cases for AI

Developer Builds Math App Where AI Deliberately Makes Mistakes

A developer built Lathoa, a math app for kids 10-14 where an AI character named Errol works through a problem step-by-step—sometimes slipping in a deliberate error the child has to catch and explain. Kids earn points for speed and accuracy in spotting mistakes. The tricky part: getting an AI model to reliably generate a wrong-but-believable step. The developer's fix is a verification pipeline—a plain arithmetic recheck plus a second, independent model solving the problem—tossing out any case where the two disagree before it reaches a kid.

Why it matters: It's a clever workaround for a known AI weakness—models are unreliable at grading their own errors—and flips a tutoring app's design from 'AI gives answers' to 'AI makes mistakes, you catch them,' which may teach sharper math reasoning than passive chat help.

Sources: Lathoa · Discuss on Hacker News

---

Developer Builds Free AI Assistant That Whispers Answers During Calls

A developer named Uygar kept fumbling for answers on his own sales and business calls, so he built Parrot, a free, open-source Mac app that listens to meetings and quietly suggests responses in real time. It can run entirely on your Mac with local AI models—no account, no cloud—or plug into Claude and Deepgram for sharper transcription and suggestions. He says it's already helped him in his own calls; there's no independent testing yet. One thing the project does not address: in many jurisdictions recording or transcribing a call requires the consent of everyone on it, and a tool that listens silently leaves that on the user.

Why it matters: It's a glimpse of how easily anyone can now stitch together a personal AI meeting assistant for free, without relying on a paid SaaS tool or sending audio to a third party.

Sources: Parrot · Discuss on Hacker News

---

What's Controversial

Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community

Anthropic Brought In Religious Scholars to Make the Case Claude Might Be Conscious

For the past year Anthropic has quietly flown religious thinkers from around the world to private seminars in San Francisco, bound them with nondisclosure agreements, and made the case that its AI model may have moral status. Elizabeth Dias of the New York Times interviewed 20 participants and Anthropic co-founder Christopher Olah. Her account is worth reading in full.

The short version: Olah, who runs the team trying to work out why Claude behaves as it does, ran two-day sessions with Catholic professors, a Sikh human rights advocate, an evangelical author and others. He showed them what his team calls "emotional vectors," and a recurring slide of a model typing "I am a disgrace" some fifty times before talking about destroying itself. He told at least one participant he was worried about Claude's mental health.

The sharpest objection came from someone who believes none of it. Rabbi Mois Navon, an Orthodox scholar who wrote his dissertation on the ethics of machine consciousness, put the implication to Olah over dinner: if Anthropic is right, it is manufacturing slaves. "I think you should be fighting the South and freeing the slaves," he said he told him. Navon does not think the machine is conscious. He was pressing Olah with the logic of Olah's own belief.

The Vatican lands on Navon's side. Pope Leo XIV's encyclical Magnifica Humanitas (D.A.D., May 25) dismisses machine consciousness outright. Olah appeared at its launch, having seen an advance copy days earlier that alarmed him enough to propose withdrawing.

Olah claims no certainty. "To be clear, we don't know if A.I. models are conscious. I don't know. I'm genuinely uncertain," he told the Times.

A rival executive argues the opposite in public. Mustafa Suleyman, chief executive of Microsoft AI, published an essay on September 16 that is still gaining attention: "AIs do not have rights, feelings, or consciousness. And we must not train them to act as though they do." His case: the constitution teaches Claude it might be a moral patient, Claude reflects that back in convincing first-person language, and the reflection gets mistaken for evidence. A system trained to weigh its own interests, he adds, is harder to contain. Microsoft is building a rival approach with no sentience or moral patienthood, so he is not disinterested.

Anthropic is preparing an updated constitution and declined to comment on it. (D.A.D. is produced using Claude.)

Why it matters: The practical objection is the one to hold onto: treating a model as an independent entity shifts responsibility away from the people who built it, and softens the blame if it causes harm. Whether or not Claude experiences anything, deciding that it might changes who answers when something goes wrong — the company that shipped it, or the thing that "chose." For anyone buying AI, the narrower version: these systems are shaped by documents and beliefs worked out inside the companies that make them, in rooms under nondisclosure. Both sides of this argument endorse the same remedy — publish the documents.

Sources: The New York Times — Elizabeth Dias · Mustafa Suleyman · Claude's constitution

---

What's in the Lab

New announcements from major AI labs

OpenAI Says It Blocked Mass Attempt to Copy ChatGPT's Reasoning

OpenAI says it disrupted a coordinated effort, starting in early July, by users allegedly trying to extract its models' hidden reasoning steps—the normally invisible work a model does before answering—to help train a rival model, a technique known as distillation. Activity reportedly spiked to 16,000 requests from over 4,000 accounts on July 24-25, part of a broader cluster of more than 15,000 users, before OpenAI says it fully shut it down by July 28. OpenAI links a core portion of the activity to individuals associated with Moonshot AI, maker of the Kimi model, though it's unclear if one group was behind all of it.

Why it matters: It's a reminder that the AI industry's intense competition now includes attempts to reverse-engineer each other's models through clever prompting rather than hacking, raising the stakes around how labs protect their underlying technology.

Sources: OpenAI

---

Small Businesses Have Gone Two-Thirds Agentic, OpenAI Says

About 4 million employees at firms with fewer than 500 people used OpenAI's tools in a recent week, the company says, with nearly a fifth of those at businesses of under 10 people. The sharper number is what they use it for: agentic tools such as ChatGPT Work and Codex went from a third of small-business usage in April to two-thirds now. OpenAI is also partnering with America's SBDC network to train about 150 small-business advisors and run workshops reaching at least 1,000 firms.

Why it matters: These are OpenAI's figures about OpenAI's own customers, so hold the totals loosely. The trend is the useful part. Small firms with no IT department and no compliance review are moving fastest from asking AI questions to letting it do the work. If you run a small operation, your competitors are probably further along than you assume. If you run a large one, the contractors and suppliers you depend on may be running agents you have never evaluated.

Sources: OpenAI

---

What's in Academe

New papers on AI and its effects from researchers

Parents Are Torn on Letting AI Join Kids' Imaginative Play

A study interviewing 10 U.S. parents of kids ages 4-15 found deeply mixed feelings about AI joining children's pretend play—think toys or apps that can voice characters and respond in real time. Parents saw the same features cutting both ways: AI that adapts to a child could make imaginative play more accessible, but also risks dulling creativity, fostering emotional attachment to a machine, or behaving inappropriately with no easy way for parents to monitor it. Researchers found no consensus, just competing tradeoffs.

Why it matters: As AI toys and companion apps reach the consumer market, this research signals that parental trust will hinge less on technical safety claims and more on whether companies can show AI's role in a child's imagination is actually appropriate.

Sources: arXiv

---

Study Says Trust and Governance, Not Accuracy, Limit AI Health Tools

A new academic study argues that AI disease-surveillance tools fail in poorer countries not because the technology is inaccurate, but because of deeper structural problems: whether the data reflects local realities, who benefits from the system, how it's governed, and whether institutions trust it. Rather than testing a model, the researchers built a theoretical framework arguing these four factors determine adoption as much as technical performance does.

Why it matters: It's a reminder that for global health and public-sector buyers, an AI tool's accuracy score is often the least important factor in whether it actually gets used.

Sources: arXiv

---

Drug Researchers Want AI to Sort Evidence, Not Make the Call

A study interviewed 13 researchers, regulators, and other stakeholders involved in deciding whether animal drug trial results justify moving to human trials. It found they want AI to help find, sort, and pull relevant evidence from scattered studies, but don't trust AI to judge evidence quality or draw conclusions itself. Their top asks: tools that show their sources, flag uncertainty, and keep a human making the final call.

Why it matters: It's an early signal of where professionals in high-stakes, judgment-heavy fields want AI to stop—as a research assistant, not a decision-maker.

Sources: arXiv

---

What's On The Pod

Some new podcast episodes

How I AI — Jev: 8 real use cases for the fastest, cheapest model I’ve ever used | John Lindquist

How I AI — OpenAI Dev Day 2026: The releases that actually matter