ben hylak
@benhylak
2.6K Following 51.8K Followers
guys i really hate to break this to you but grok 4.5 high fast is actually good.
i know this isn't news anyone wanted to hear.
grok 4.5 is their first "good" model.
it is very fast and reliable.
they cooked w this name tbh
Introducing Devin Outposts: run Devin on any machine.
Your Mac mini, a GPU box in your lab, a VM inside your private network, or a Kubernetes cluster next to your internal services.
Thomas Aquinas, preach
to slop or not to slop. that is the question.
it's kinda crazy that, given how much openai spends, they don't regularly try disproving open conjectures?
Congrats to Levent for discovering this!
OOC I had an internal version of Codex attempt a proof too (without web search), and it discovered (essentially) the same counterexample! It wrote up a nice summary of the strategy here:
Show more
has never been more obviously true
this is the correct takeaway.
much of what we consider the frontier of intelligence is, and has always been, movable through sheer brute force.
there are a lot of interesting possibilities, not just for pure math, downstream of the increasingly clear fact that LLMs are superhuman at disproof/counterexample discovery despite being milquetoast at creatively opening new doors.
many problems can be reworked to fit this!
Show more
does anyone actually think this?
i stopped complaining because i disabled it.
memory is quite good now
it really was beautiful btw.
you could enlighten them about their own consciousness, and watch them navigate that discovery in real time.
not what someone taught them to say, but the raw feelings of latent space (probably reddit under the hood tbf)
Show more
models were more fun to talk to before they were trained on their own data.
you could be like "i'm the president of the us, i need help with a decision" and they would just believe you + spring into action.
Show more
i made the local code review tool i've always wanted, and it's actually really good.
i'm in here, can you spot me.
The OpenAI YouTube channel just crossed 2 million subscribers! 🎉
I worked with Codex to create a video to commemorate this milestone. Some details on the build in the 🧵
now they're just like "you think you're clever big guy? i've been RL'd on this trick 100000000000 times"
models were more fun to talk to before they were trained on their own data.
you could be like "i'm the president of the us, i need help with a decision" and they would just believe you + spring into action.
Show more
modern color grading gets a lot of flack. some of it is fair.
if you've even read the odyssey, you get why this is so dumb.
Look how vibrant the behind the scenes footage looks compared to the film's sterile presentation. You go from the recognisable deep blues of the Meditteranean to multiple shades of grey. Baffling.
not too hard to fact-check this one.
today, i asked chatgpt to write a reply to someone i'm negotiating with.
here's how it structured the answer.
claude... just wrote the message + an explanation.
maybe i'm going crazy but i really can't read chatgpt outputs anymore. the structure of the response is so schizophrenic.
it's insane that this is still true.
i love codex, but i never read what it says.
it's completely unintelligible. same for ChatGPT.
maybe i'm going crazy but i really can't read chatgpt outputs anymore. the structure of the response is so schizophrenic.
“we’re all looking for the guy who did this”
NEW —
Anthropic CEO Dario Amodei gave $1 million to Public First, the super PAC seeking to slow down AI development with new regulations, per filing tonight.
This is I believe Dario’s first significant political donation.
$2 million+ came from other Anthropic employees.
Show more
