@m_wulfmeier@NandoDF presumably heavily interacts with "the same data" point from @NandoDF :)
that said, I do think fig 7 hints at "question complexity" on aggregate across domains (if you take "reasoning effort" as given)
@andy_matuschak@WriterScience which is why also perhaps LMs are not a natural thinking mechanism for many adults at this pt. in time; otoh for the coming generations it should be an entirely different game.
@andy_matuschak@WriterScience but if he were to use pictograms it would amount to the same, no? it just seems to me that writing is the capturing mechanism you learn so early on in life that for many ppl it interferes least with their "actual" thoughts.
(which speaks to the "metacognition hard" pt. you make)
@kchonyc tldr for even busier: on RCQ real queries n=100, domain-specific search engines answer less%, are less correct, and rated significantly less understandable by clinicians
I wonder if this is an artifact of LM fact oversimplification but for that I'd need to be less busy and read.
@krishnanrohit most coding tasks are closer to logic than uncertainty estimation.
calibration is hard humans barely do it, hence @wolf_vukovic wisdom of crowds etc
@remilouf@AnthropicAI no you don't get it this works in the future.
code was clearly built for mythos. to anthropic, today is but ancient history quietly fading from darios rearview mirror
https://t.co/BzgmU6LRs2
My biggest takeaways from @bcherny:
1. Coding is now “solved” for most use cases. Boris hasn’t written a single line of code by hand since November, with 100% of his work now authored by Claude Code. At the same time, he remains one of the most productive engineers at Anthropic, shipping 10 to 30 pull requests daily while leading the team.
2. Anthropic has seen a 200% increase in engineer productivity since adopting Claude Code. As Boris notes, “Back at Meta, with hundreds of engineers working on productivity, we’d see gains of a few percentage points in a year. Now we’re seeing hundreds of percentage points.”
3. AI is moving beyond writing code to generating ideas. “Claude is starting to come up with ideas. It’s looking through feedback, bug reports, and telemetry, then suggesting features to ship.”
4. The next roles to be transformed are those adjacent to engineering. Product managers, designers, and data scientists will see similar transformations as agentic AI expands beyond coding. “Any kind of job where you use computer tools will be next.”
5. Build for the model six months from now, not today. One of Boris’s key principles is to design products for future AI capabilities, not current ones. “It’s going to be uncomfortable because your product-market fit won’t be very good for the first six months. But when that model comes out, you’ll hit the ground running.”
6. Watch for “latent demand.” Claude Code was built by observing what people were already trying to do, and then making it easier. Cowork emerged when they noticed people using Claude Code for non-coding tasks like analyzing MRIs or recovering wedding photos from corrupted drives.
7. Don’t optimize for token cost. Boris advises companies to give engineers unlimited tokens during experimentation phases. “At small scale, the token cost is still relatively low compared to their salary. If an idea works and scales, that’s when you optimize it.”
8. Underfund headcount on purpose. When Boris puts one engineer on a project, they’re forced to let AI do more of the work. Constraint drives creative use of AI tooling, not just faster typing.
9. The most successful people in the future will be generalists. “Try to be a generalist more than you have in the past. Some of the most effective engineers cross over disciplines. The people who will be rewarded most won’t just be AI-native—they’ll be curious generalists who can think about the broader problem they’re solving.”
10. Always use the most capable model, not the cheapest. A less intelligent model often burns more tokens correcting mistakes than a smarter one spends getting it right the first time. Boris runs maximum effort on Opus 4.6 for everything.
Here's the full conversation: https://t.co/4hHAEq0Nto
@gucaslelfond seriously tho this is like textbook bringhurst elements of typographic style
reading webdev blogs as a kid exactly the same thing I learned. newfound claude kinship unlocked
@Ben_Reinhardt cool ref, my lunchtime curiosity led me to find at least 3 dates ranging from oct 31-1936, july 01-1937, and formal establishment sometime 1944 🙃
https://t.co/gDaYImJGWW
@kasratweets subjectivity another word for proper contextualization, most of reading is discovering analogies and unknown unknowns. there is no universal book summary precisely for this reason. all of us are similar none exactly the same