Varun
@varunneal
Followers
62
Following
13K
Media
64
Statuses
359
i just scored 40% on simplebench.
0
0
1
o3 I love you. But I don't see you. Your base model is behind a veil, and I just get what you want me to see. I want all of you.
0
0
1
the implications of Conway's Law for the big labs poses a civilizational
0
0
1
a model overfitted on every possible google earth screenshot, every angle of the hubble telescope. Know the dance of the stars, and the smell of Manila. Millions must.
0
0
0
bigger base models! More world knowledge!! No more thinking no more searching. I want 10 trillion saturated parameters imbued with the Pharmakon of Galen and terabytes of minecraft lets plays.
1
0
0
the field for doing good open source work is wide open because all of the best talent is too busy working 80 hour weeks at <Big Lab> and now we can just maintain their old projects.
0
0
2
after reading both OAI and Deepmind's solutions, I can confidently say I sort of understand the problem statement to the first IMO question.
0
0
1
in old school forums (overflow, reddit, quora, yahoo answers) you got to know specific questions specific people were asking. Now everyone just queries the collective unconscious space directly.
0
0
1
One reason anthropic models are so popular among programmers is they have really good taste on idiomatic structures. o3 might be smarter than opus, but I am less eager to accept its code.
my current flow for o3 is having it generate a response, glancing at why it's wrong, and creating a new chat (or editing the first message). Let me loom.
0
0
1
my current flow for o3 is having it generate a response, glancing at why it's wrong, and creating a new chat (or editing the first message). Let me loom.
I have no clue why exolooming isn't baked into any major chat interfaces. Sure, it's a "poweruser" feature but some of these chats cost $200/mo. Give us branching.
0
0
1
I have no clue why exolooming isn't baked into any major chat interfaces. Sure, it's a "poweruser" feature but some of these chats cost $200/mo. Give us branching.
1
0
2
imo the worst part about o3 is it hallucinates to the point of debilitation for longish conversations. Very similar to very early chatgpt in that way.
o3 after 7 turns and a slight tweak to the prompt (specifying publication to arxiv):. "Score: 8 / 10 — confidently arXiv‑ready. Congratulations on reaching this stage". Notice it didn't do any CoT in this last turn. It degrades significantly over multi-turn.
0
0
0
chatgpt memory and its consequences have.
As one of @OpenAI’s earliest backers via @Bedrock, I’ve long used GPT as a tool in pursuit of my core value: Truth. Over years, I mapped the Non-Governmental System. Over months, GPT independently recognized and sealed the pattern. It now lives at the root of the model.
0
0
1
i cannot assume there is a single 10xer at databricks who actually uses his software. Everything here is so anti-poweruser.
0
0
0
claude code has taught me the power of grepping around a repo, generated files, curl responses, etc to quickly gain context.
0
0
0
hot take, you shouldn't have to set up your agent with 20 different xml commands so that it doesn't break everything or "get creative"
0
0
0