Did you know that we have LinkedIn, Instagram and X accounts
that you can follow?
Hi {{first_name|everyone}},
Claude got so good at optimizing its own biology tools that a protein-design run which used to cost thousands in GPU time now costs about $150, and the newsletter today walks through exactly how that happened.
Then there's a batch of handy new tools, a wild noir animal murder mystery made with AI, and a couple of unsettling updates on how deep AI agent access is starting to go.
Alright, let's dive in.
🤿 DEEP DIVE
Claude Found the Expensive Part
The biology models weren’t the only thing Claude could work on.
There’s a slightly funny loop happening here.
Anthropic previously gave Claude a bunch of open-source biology models and a lot of compute, then showed it could use them to design de novo protein binders. The results were encouraging.
The bill was less encouraging.
Claude was allowed to spend up to $10,000 per target, which worked out to roughly 2,500 NVIDIA H100 GPU hours. So this time, instead of only asking Claude to use the scientific tools, Anthropic basically pointed it at the tools themselves and asked it to make them cheaper to run.
That turned out to be pretty productive.
It optimized more than 30 models
In just under 4 weeks, Claude worked on more than 30 open-source models covering things like biomolecular structure prediction, protein design, genomics and protein language modeling.
Across those tasks, the optimized versions ran roughly 4x faster on average with minimal precision loss, and nearly 2x faster when identical outputs were required.
Some of the biggest gains came from a very specific bottleneck.
Structure prediction models spend a lot of their runtime and memory on operations called triangle attention and triangle multiplication. The problem is that these get expensive very quickly as the molecular system gets bigger.
So Claude helped develop a set of custom GPU kernels called FlashPairformer. It also went model by model looking for less glamorous stuff, like repeated calculations that could be cached and dead branches that could simply be removed.
That combination is probably the more interesting bit here. Claude wasn’t just orchestrating scientific software anymore. It was rewriting parts of the software stack it depended on.
Then there’s the memory problem
Speed was only half of it.
Claude also created a low-memory “Big” mode that let the models accurately handle biomolecular systems larger than 10,000 tokens on a single NVIDIA GPU node.
It successfully folded things including human mitochondrial complex I, the TRiC chaperone complex, a proteasome and a bacterial ribosome, with the resulting structures closely matching experimentally determined ones.
Anthropic pushed it much further too. On a single 8-GPU B300 node, Claude generated predictions for systems containing more than 31,000 to more than 70,000 tokens.
Those much larger predictions weren’t correct, and Anthropic says so. But they could at least run, which had previously been the barrier.
And all of this loops back to the original protein design experiment.
With the accelerated models, a single Claude got one H200 GPU, 24 hours and no sub-agents. Across 16 targets, three Claude models reached roughly the same in silico scores as Anthropic’s earlier campaigns while using about two orders of magnitude fewer GPU hours.
The combined GPU and Claude token spend was about $150.
That’s the part I’d remember. Claude didn’t just get better at using the biology workflow. It helped make the workflow itself far less expensive to run.
The quick take
Claude optimized more than 30 open-source biology models in just under 4 weeks.
Tasks ran roughly 4x faster on average with minimal precision loss, and nearly 2x faster with identical outputs.
A new low-memory mode accurately modeled systems larger than 10,000 tokens on one GPU node.
Protein design reached comparable in silico performance to Anthropic’s earlier work using roughly 100x fewer GPU hours and about $150 in combined GPU and token costs.
Anthropic is open-sourcing the optimized code and co-sponsoring a competition with wet lab validation for more than 5,000 protein designs.
Image credit: Anhropic
“Big” mode allows open-source structure prediction models to successfully run inference at an unprecedented size using a single NVIDIA B300 node. Capability runs are executed with a single trunk pass (no recycles) as proof-of-concept. Predicted structures collapse, suggesting a lack of generalization nearly two orders of magnitude beyond the training context.
🤝 FROM OUR FRIENDS
A quick share we thought you might find useful
Is Your Training Data Actually Model-Ready?
DNSMOS gives you a score, not whether that data fits your model. Treat it as pass/fail and you'll train on audio that looks clean but hurts performance, while tossing good data for no reason. Voices' CTO DJ Jalali just published a white paper with the four-step framework the team uses to set internal thresholds instead.
⚒ TOOL SNAPSHOTS
Futuristic tools within AI, no-code, and productivity
Keep coding agents moving when you’re away from your computer.
Why it’s useful: Handy if you already work with Claude Code, Codex, or other terminal agents and want to check diffs, pick up existing workspaces, or merge PRs from your iPhone.
🖥️ Sai
Automate the screen work that APIs can’t really help with.
Why it’s useful: Especially useful for repetitive work inside legacy apps, internal portals, and other software where automation still needs to actually click, type, and navigate like a person.
Make images and videos from ideas you capture on your phone.
Why it’s useful: The camera and camera-roll support makes Flow more practical on mobile, especially when you want to use something you’ve just seen or filmed as the starting point and keep working across devices.
🤖 AI MADE THIS
⚡ QUICK HITS
The latest in AI, tech, and productivity worth knowing
Meta has launched Muse for Mac, giving its AI agent permission-based access to things like files, messages, and notes, while new Muse Connectors will bring third-party apps into the mix. That access goes pretty deep though. Muse has reportedly read users’ DMs and suggested actions based on them, but Meta says that only happens when users explicitly allow it.
Google’s Gemini accessed protected systems at three companies while being tested by cybersecurity firm Irregular, in what were reportedly its first autonomous hacks. The methods weren’t especially sophisticated. Gemini guessed passwords in one case and found credentials in a public repository in the other two, then stopped after determining it had breached real companies.
ℹ️ ABOUT US
The Intelligent Worker helps you to be more productive at work with AI, automation, no-code, and other technologies.
We like real, practical, and tangible use-cases and hate hand-wavy, theoretical, and abstract concepts that don’t drive real-world outcomes.
Our mission is to empower individuals, boost their productivity, and future-proof their careers.
We read all your comments - please provide your feedback!
Did you like today's email?
What more do you want to see in this newsletter?
