Read documents. Not too many. Mostly PDFs.
Our parents read and wrote documents. Having the world's knowledge an internet connection away doesn't change this.
@JFPuget Maybe the same reason a) cats play with mice or b) people enjoy watching cats play with mice? I found Tim Gowers' explanation of this helpful in my own work.
Gemini Flash 3.5 is such a disappointing model.
It's intelligence and speed is awesome. Absolutely amazing.
But it's been trained to max evals, not to be helpful to humans.
It goes off and does random crap "for me" rather than just doing what I asked.
@HanchungLee@Flomerboy If AI+competent IC produces 10x ICs then base managers # reports goes 1/10, usually ~1. 10x ICs don't need 1 person managers. That strips one layer of management. Manager turned IC is another path. That removes her 10 IC reports. Where's the residual management work? Coffee?
BrowseComp-Plus, perhaps the hardest popular deep research task, is now solved at nearly 90%...
... and all it took was a 150M model ✨
Thrilled to announce that Reason-ModernColBERT did it again and outperform all models (including models 54× bigger) on all metrics
financial literacy psa: payment processing volume is 1000x larger than gdp.
so, yes, ramp data is used only because it’s available. and is tiny enough that it can’t be extrapolated.
@yoavgo I think it's just giving a name to what you already knew . I just meant that the amount useful new software and slop will both be proportional the number of new programs written by AI . Numerate people intuit this but it confuses some others.
@lemire The lowest energy path to Founder Mode is to bypass the clerical middle-management layer with AI to speak directly to the makers/developers. Systems eventually move to the lowest energy states.
The degree to which AI research at the big labs has almost entirely been reduced to hill climbing is actually an aberration and not reflective of the rest of science at all. Ironically this means AI research is probably the easiest branch of research to automate.
I think one of the conclusions we should draw from the tremendous success of LLMs is how much of human knowledge and society exists at very low levels of Kolmogorov complexity.
We are entering an era where the minimal representation of a human cultural artifact... (1/12)
Life in Chrome: I try to drag a tab somewhere, and if I let go of it at the wrong time, it disappears! Where has it gone? It's now part of a split screen with some other tab. Eventually I find and separate them and try again. I've never seen a feature that felt more like a bug.
While this is true but I'm not completely convinced CS will stay a relevant category (or much tinier, almost back to pre-1990s). With effective agents, computation is just too ubiquituous.
@remilouf That's the key change to the economy; not exisiting companies becoming more capable at building software. Big companies don't have this bottleneck, tiny ones do. I expect a vast increase the startups that can move from interesting PoC to polished product? How will this play out?
Cannot wait for teams that build their custom JIRA/Workday replacement/custom CRM to one day turn around and ask:
“Why do we have so much internal software that is buggy / has poor UX + we need to maitain?”
Seen this movie well before AI, when Uber built uChat (custom Slack)
We just ruined all of science because we can get a measure of performance to improve by (checks notes) 2% without a scientist within the loop (but truly expensive people making the loop). Comparison to standard old fashioned hyperparameter optimization not necessary.
This one from Hacker News .. profound, loving, nerd-perfect obituary written entirely in the formal language that Sir Tony Hoare himself invented: CSP (Communicating Sequential Processes).
SIR_TONY_HOARE = μX • (think → create → give → X)
-- process ran from 1934 to 2026 -- terminated with SKIP -- no deadlock detected -- all assertions satisfied -- trace: ⟨ quicksort, hoare_logic, csp, monitors, -- dining_philosophers, knighthood, turing_award, -- billion_dollar_apology, structured_programming, -- unifying_theories, ... ⟩ -- trace length: ∞ The channel is closed. The process has terminated. The algebra endures.