Charles Weill

agentmaxxing | ex-{ @xAI, YC W23, @GoogleAI, @Cornell }

[Charles Weill @weill]

@RealJoshHoward @CreatorML_ From what I’ve seen, frontier models like Opus and GPT don’t do much better whether you include the thumbnails or not as inputs to the prediction. Could be that they need your signal processing first

[Charles Weill @weill]

I’m considering open-sourcing an eval based on 4 years of trying to build AI that answers: ‘What should my YouTube audience watch next?’

At @CreatorML_ we built view predictors, but we always lacked a clean public benchmark.

Curious: how much better are Claude Opus/GPT-5/etc. out-of-the-box at this vs specialized models? Planning to test it properly.

[Charles Weill @weill]

Let me know if you want to contribute data or compute to get some baselines from SOTA models.

[Charles Weill @weill]

Key question: Given channel history + new video idea (title/thumbnail/desc), how well can models predict relative performance (views, CTR, retention)?

Predicting “is 1 of 10” might be good starting point as opposed to raw views.

[Charles Weill @weill]

@_jonatasantos Thoughts on using beads for per project memory and issue tracking?

[Charles Weill @weill]

Seems we have converged to the same setup.

[Jonata Santos @_jonatasantos]

how to build anything rn:

- get a hetzner, do, or hostinger vps
- host hermes on it
- add gbrain or implement your own memory vault using qmd + sql
- set up hermes with codex auth -> gpt-5.5 / no reasoning / fast mode
- install orca on your macbook and phone with tailscale to have a nice ide to work on both
- before starting any work, ask hermes to conduct deep research on the subject and save it to gbrain as source material for the project
- use the `/grill-me` skill or a similar prompt to uncover as many unknowns as possible. save results to memory too
- define/write clear evals for every project to determine whether a run was successful
- have hermes iterate over the project until all evals pass, saving all learnings to the vault along the way
- whenever it gets stuck, use memory + a new research or `/grill-me` session to unblock it

rinse and repeat until the work is done. pay attention to the process. develop a feeling for how long tasks should take and do not be afraid to stop a model mid session to ask for status and why it's taking so long.

[Charles Weill @weill]

@dexhorthy Yep. Codex 5.5 is still the best coder. But sometimes 4.8 surprises me in good ways. I still wouldn’t rely on Claude for anything mission critical

[Charles Weill @weill]

Silly codex

[Charles Weill @weill]

"...you're right, ..."
"... objective fact ..."
"... honest ..."

[Charles Weill @weill]

@GoodisonJason Einstein said, "The secret to creativity is knowing how to hide your sources." 🧠

[Charles Weill @weill]

“People who are really serious about software should make their own hardware” —Alan Kay

People who are serious about AI should use specialized chips for inference.

[Jason Goodison @GoodisonJason]

We raised $15m to build the ASICs-first inference cloud.

We're betting big on alternatives to GPUs, and the result is that we are already 5-8x faster on most models.

Read more about General Compute on Tech Crunch!

@FPuklowski @fastinference

https://t.co/xQyZmfVJPN

[Charles Weill @weill]

@GoodisonJason @FPuklowski @fastinference Oversubscribed! Congrats

[Charles Weill @weill]

@yacineMTB Claude is a great coordinator though. And party planner for all the codexes that won’t dance

[Charles Weill @weill]

@yacineMTB Exactly. I only let Claude write code when I run out of Codex subscription

[Charles Weill @weill]

$400/mo Codex plan with 50x tokens when? @openai

I’m all out of two separate 20x subscriptions.

[Charles Weill @weill]

> switch to Opus 4.8
> immediately run out of quota

💀

[Charles Weill @weill]

@doodlestein Same reason I made my own tmux to better fit my own workflow

[Charles Weill @weill]

tmux

iykyk

[Charles Weill @weill]

https://t.co/514yvdZJHi

[Charles Weill @weill]

I'm playing around with my own agent swarm harness/flywheel to rebuild DeepMind's AlphaZero.

I'm learning so much.