1- running agents for longer dones not increase success rate to 100% : a dumb model will never solve hard tasks no matter how long it runs
(Same as in an IQ test actually, to use your analogy: spending even days on a task you don't understand won't help you solve it)
2- running
@aymericroucher
-
AI Agents’ Limitations in Solving Hard Tasks
By
–
-
Doubling AI Task Autonomy and LLM Training Challenges
By
–
According to METR, the length of tasks that AI can solve ("time-horizon of autonomy") is doubling every 7 months But that leaves many questions unanswered ▸ What abilities in LLMs did it take to increase autonomy that much?
▸ How will we train LLMs to keep this progress -
OpenAI’s Projected Training Costs and Scaling Forecasts
By
–
OpenAI's own projections for burn shocked many :
– $35 billion in 2027
– $45 billion in 2028
(most will go into model training) Yet even numbers were already forecasted one year ago by @leopoldasch in Situational Awareness (table below) Keep calm and scale on. -
LLMs and the impact on entry-level job tasks
By
–
There could be other effects! For me it's quite intuitive that entry-level jobs are more affected, because it's the jobs whose tasks are generally lower-level, so more accessible to LLMs
-
Analyzing the Current Economic Impact of AI Infrastructure
By
–
Spot on article by @DKThom
:
"All this talk about AI as the technology of the future—will it cure cancer in 2030? or, destroy the world in 2027? or accomplish both, maybe within the same month?—can evade the question of what AI is doing to the economy right now. AI infrastructure -
Optimizing AI Coding Workflows with System Instructions
By
–
Daily reminder : ~/.codex/instructions.md or ~/.claude/CLAUDE.md make a huge difference! My instructions:
– Don't catch errors, I prefer to raise them and fix them myself
– Go simple
– For GPT-5 in codex: don't abbreviate names (model has a tendency to make confuse -
Evaluating Gemini 2.5 Flash on ARC-AGI benchmarks
By
–
I've tested if nano-banana / Gemini-2.5-flash-image beat ARC-AGI – it's quite far. Btw bravo to the ARC_AGI team, the delta between easiness of problems for humans vs difficulty for LLMs is just
-
Claude Code’s unique prompt handling feature explained
By
–
One cool thing that Claude Code has is the ability to type guidance as the model goes, and when you hit enter it's put in a wait stage and submitted to the prompt after the next tool call.
I didn't see this in Codex. -
GPT-5 Achieves New Benchmark on Training Info Efficiency
By
–
GPT-5 just beat a new benchmark: the "model training info / total system card length" is now approaching 0 !