- AI-written genomes produced 16 working bacteriophages in a Stanford and Arc Institute study, the first functional viruses designed end to end by a model.
- A companion editorial from Johns Hopkins biosecurity researchers says the screening systems meant to catch dangerous DNA orders weren't built for sequences a machine invented.
- Microsoft is collapsing Copilot chat, GitHub Copilot, Copilot Cowork, and its Autopilot agents into one app behind a single login.
- ByteDance is reportedly pre-training a model with up to 10 trillion parameters, one of the largest runs any lab has attempted.
- If today's theme is AI leaving the screen, our piece on where generative AI turns into functional AI is the background reading.
For three years the argument about AI has mostly been an argument about text on a screen. Today's news is about what happens when the output lands somewhere physical. A model wrote a genome and the genome worked. Another company is spending billions on the buildings and cables that make any of it run. A third is merging four products into one place because people have stopped wanting a separate app per capability.
Underneath all of it sits the same gap. The systems we built to check things, biosecurity screening, procurement review, safety testing, assumed a human made the thing and made it slowly.
The Front Page: A Model Wrote a Virus and the Virus Worked
Researchers at Stanford and the Arc Institute published a study in Science on August 6 describing the first functional viruses designed by AI. Using two genome language models, Evo 1 and Evo 2, they generated hundreds of synthetic genomes for phiX174, a bacteriophage that infects E. coli. A bacteriophage is a virus that attacks bacteria and nothing else, and it's the workhorse of a whole field of medicine that uses them to kill infections antibiotics can't touch. Sixteen of the AI-designed phages worked when synthesized and tested. Three outperformed the natural strain. Several got past bacterial defenses the original couldn't. BetaNews has an accessible summary and the University of Reading collected outside expert reaction.
The team took the obvious precaution: sequences from viruses that infect humans, animals, or plants were held out of the training data on purpose. Credit where it's due, that's a real design choice and not a press-release flourish. But the biosecurity specialists at Johns Hopkins who wrote the companion editorial in the same issue land somewhere less comfortable. Their point isn't that this study was reckless. It's that the governance for the capability the study just demonstrated doesn't exist yet.
What it means: The screening layer here is a set of databases that DNA synthesis companies check orders against, looking for sequences that resemble known dangerous ones. That whole approach assumes the thing you're looking for has a relative in the reference set, because until now everything came from evolution or from copying evolution. A model that writes genomes unlike anything in nature breaks the resemblance test. Nobody has to be malicious for that to be a problem, and nobody needs a chatbot jailbreak either. This is the clearest example yet of a pattern worth internalizing: AI doesn't only speed up existing work, it produces artifacts our verification systems were never designed to recognize.
Releases & Features
Microsoft is building one Copilot instead of five. Satya Nadella confirmed that Copilot chat, GitHub Copilot, Copilot Cowork, and the Autopilot agent layer are merging into a single app this year, joined by a shared identity graph so one login covers GitHub and Microsoft 365. Techweez has the shape of it. In practice that means asking a question, moving into code with the conversation still attached, then handing the whole thing to an agent without re-explaining anything.
Grok Voice got noticeably quicker. As of August 5, xAI's Think Fast 2.0 is the default behind grok-voice-latest, cutting time-to-first-audio to about 0.70 seconds from 1.25, priced at $0.08 per minute of speech-to-speech. Those are xAI's own figures. Under a second is roughly the point where a voice interface stops feeling like a walkie-talkie and starts feeling like a conversation.
What it means: Both moves are about friction rather than intelligence. Microsoft's bet is that the thing slowing adoption isn't model quality, it's four logins and no shared memory between tools. xAI's bet is that half a second of silence is what makes people give up on voice. Neither company claimed a smarter model this week, and that's the tell: the competition has moved to whether the thing fits into an actual working day.
In the Lab
Reuters reports that ByteDance is pre-training a model with up to 10 trillion parameters. Parameters are the adjustable numbers a model learns during training, the closest thing it has to knobs, and 10 trillion of them would be roughly three times the size of Moonshot's Kimi K3 and larger than outside estimates for Anthropic's Mythos 5. Pre-training is the long, expensive first phase where a model reads its training data; it typically runs three to six months before anyone fine-tunes or ships anything. Founder Zhang Yiming has reportedly told teams to chase real capability rather than distilling a bigger model down into a cheaper one. Details here.
What it means: This is a contrarian bet, and worth watching precisely because so much of the last year argued the other way. The industry's fashionable position is that scale has plateaued and the wins now come from training tricks, tool use, and smaller specialized models. ByteDance is spending an enormous amount of money on the opposite view. If it pays off, expect the pendulum to swing back hard. If it doesn't, we'll all pretend we knew.
The Oversight Desk
AI music company Suno says it will add watermarking and audio fingerprinting so tracks made on its platform can be identified, and is working with distributors on ways to flag fraudulent or spam uploads. The announcement arrives while Suno is in litigation and negotiation with major music companies, which is worth holding in view when reading it.
What it means: Provenance is becoming table stakes, and not because companies love it. The EU's transparency rules already require machine-readable marking of synthetic output, streaming platforms are drowning in mass-uploaded tracks, and labels want a way to tell what came from where. The honest caveat is that watermarks are fragile: re-encode a file, run it through effects, and the mark can degrade or vanish. Treat any provenance claim as a signal rather than proof, whether you're a musician checking a track or a company checking a supplier's deliverable.
Microsoft's whole pitch today was fewer apps and one context. You can get most of that yourself by naming the handoffs you keep doing manually. Say what they are and BYOBot writes the agent.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- AMD bought Taalas. The Toronto startup, which had raised about $219 million, builds silicon that bakes parts of a model directly into the chip to cut the cost of running it. Terms undisclosed. Reuters via TechStartups.
- Google reshuffled its AI leadership. Demis Hassabis is stepping back from daily operations into a scientific role, Koray Kavukcuoglu takes over running the organization, and Jeff Dean is leaving to start a company focused on automating scientific research. The Verge via TechStartups.
- Firmus raised $2 billion. The Australian AI infrastructure company hit a valuation above $10.5 billion with backing from Nvidia, Blackstone, Coatue, and Jane Street, roughly double where it sat in April. TNGlobal.
- A Chennai robotics startup raised $5.5 million to crawl through sewers. Solinas Integrity builds robots that inspect pipelines and confined infrastructure, the kind of narrow, unglamorous automation that pays for itself immediately. The Economic Times via TechStartups.
The Bottom Line
A phage that beats its natural ancestor is a genuinely hopeful result, because antibiotic resistance kills people every day and phage therapy has been starved of good design tools for decades. The uncomfortable half is that the same capability arrived before anyone built a way to check it, which is now the recurring shape of every AI story worth reading. Watch the DNA synthesis screening standards over the next few months, because that's where this either gets addressed or quietly doesn't. And in your own much smaller corner of this, the same lesson holds: whatever you automate, decide first how you'd know it went wrong.
Frequently Asked Questions
-
It's the same basic idea as a text model, pointed at DNA instead of sentences. Train it on millions of genetic sequences and it learns which arrangements tend to follow which, well enough to write new sequences that hold together. The Stanford and Arc Institute team used two of them, Evo 1 and Evo 2, trained on sequences from about two million bacteriophages. That shift from generating words to generating working things is the one we traced in from generative to functional AI.
-
No. Parameter count measures capacity, not usefulness, and the last two years are full of smaller models matching much larger ones on the jobs people care about. ByteDance's reported 10-trillion-parameter run is a bet that raw scale still buys something cleverness can't. For most builders the question is narrower: which model does your specific task well at a price you can pay. The workflow directory is organized around tasks for that reason, and the tool landscape maps who's competing where.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
