Writing Flight Notes on AI, trading micro-structure & founder-style capital. Building in public. Opinions mine.
Joined March 2017
- Tweets3.7K
- Following2.8K
- Followers1.2K
- Likes1.7K
I shuddered in fear as a frontier lab called this beamer-looking ass a visual improvement. Do you not have money to hire artistic people? @OpenAI
> Be me, the only engineer for the entire company
> I'm awesome
> "I will just use openclaw to do everything"
> Integrate with Slack. Easy
> User requested work
> Work failed on some fringe cases
> Claw spammed 600+ messages in 30 minutes
> "No worries, I'll just fix and delete the spam messages"
> Slack rate limits API call for deletion. 1 req/s
> Shamefully have to wait 10+ minutes to delete all the spam messages my claw sent
> Awesome
> I'm still the only engineer for the entire company.
Just realized my biggest commit day was Feb 14 in Q1.
At this point, I'm just surprised my wife hasn't left me.
>be me, AI researcher
>2026, every model scores 80%+ on ARC-AGI-1
>"we're basically AGI bro"
>chollet drops ARC-AGI-3
>not even puzzles anymore
>interactive game environments
>no instructions given
>agent has to EXPLORE and FIGURE OUT what the goal even is
>mfw the model has to actually learn from experience
>like a literal child picking up a new game
>except children are better at it
>best score on the planet: 0.37%
>ZERO POINT THREE SEVEN PERCENT
>gemini 3.1 pro got the high score
>gpt 5.4: 0.26%
>opus 4.6: 0.25%
>grok: literally 0.00%
>grok didn't even show up to class
>$850K prize pool sitting there laughing
>humans solve these first try
>o3 and friends: "uhh let me think for 47 minutes"
>still can't break 1%
>"but my CoT reasoning is really good"
>yeah bro your chain of thought is a chain of cope
>turns out pattern matching on training data ≠ intelligence
>who could have predicted this
>chollet_vindicated.jpg
>the "we're 2 years from AGI" crowd real quiet rn
>mfw the benchmark name literally has "intelligence" in it
>and none of the models have any
