Bookmarks

I feel you for sure here because I'm kind of on the same track with you.

I first started to use an API and saw my money gone. That's it. This is the output. No surprise here.

Then I was thinking so this is using Claude Agent SDK under the hood, so the same thing that runs Claude Code. Why not using my Max sub?

100% of my money stress went down because now I don't need to switch to haiku to save money. I only use Sonnet because it's kind of free.

Nanoclaw is cool and I got my bot working for me and I like it, but at the same time I don't want to "hack" anything. I don't have my nanoclaw so configured at the moment.

I'll probably switch again to the API. I remember I used Openclaw at the very beginning of it and there were a lot of bugs on whatsapp so I deleted it. It was not helpful at that time.

Nanoclaw seems to be different and I very much enjoy the simplicity of it and the fact that @Gavriel_Cohen wants to make it a very useful infra.

Idk. Will see. Thanks again for poiting this out! Really helpful.

By far my biggest advice to anyone trying to adopt AI properly:

1. Pay a little bit of money to Anthropic
2. Download Claude Code
3. Open Claude Code
4. Press 'Shift-Tab' until it says 'plan mode on'
5. Open Voice Memo on your iPhone. Just talk about all the things you want to accomplish. When you think you are done, just keep talking. Make sure it is at least 10 minutes, hopefully longer
6. Send this Voice Memo to your computer
7. Download MacWhisper and use it to transcribe this voice memo. Trust me, you will want MacWhisper and will use it later a lot
8. Type into Claude Code: "I have never used you before but I talked about some things. I will paste those things in below. Please read the things and ask me any questions you need to in order to help me figure out how to use you to be awesome. Ask me lots of questions until I tell you I am done"
9. Then paste in the transcript
10. Then press enter

Then just let Claude take the wheel, and them please send me a DM if this works.

Also, if this just sounds crazy, just literally take this entire message and paste it into whatever AI you are using and say 'some weird person told me to paste this into you, I want to use it, but I don't know how. What should I do?'

I am just trying to help you get started. Curiosity and persistence are the most important things.

From 25 Years of Data to 79% Winrate on Prediction Markets

Indicators [38]

• Moving averages: MA10, MA20, MA30
• RSI (overbought/oversold)
• MACD / MACD_Signal (trend)
• Bollinger Upper/Lower (volatility)
• Volatility 10/20/30, OBV, momentum, sentiment, insider activity, etc

Humans can't do that

-> AI processes everything in seconds

Here's how it works. The purpose of the model

Not to predict the price

It answers one question: “Will there be growth tomorrow?”

Target1 = (close.shift(-1) > close).astype(int)
# 1 → BUY, 0 → HOLD

For the complete structure, you also need:

• Neural network architecture
• Monte Carlo Dropout (key point)
• Trading logic

Result:

• Win rate 59–78%
• 100+ trades per month → consistently profitable
• The human brain is incapable of analyzing 38 indicators × 30 markets × 24/7

Therefore, this trader has already made

$1.4 million in profit in 4 months

You can find a full explanation of how this works at Noisy

Original reading time: 15 min

The headline is automated hill-climbing. I'd say the deeper lesson is eval design.

The agent was not optimizing the full, noisy, expensive objective directly. It was climbing a cheap proxy that still tracked reality well enough to transfer:

• 5-minute train limit
• validation BPB as the objective over noisier ground truth metric (CORE)
• 1-GPU setup over full-scale runs on 8H100s

That’s the kind of thing people call “just engineering,” but it’s really research taste. In the agent era, designing the optimization surface may become as important as proposing the ideas.

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project.

This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.:

• It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work.
• It found that the Value Embeddings really like regularization and I wasn't applying any (oops).
• It found that my banded attention was too conservative (i forgot to tune it).
• It found that AdamW betas were all messed up.
• It tuned the weight decay schedule.
• It tuned the network initialization.

This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism.
github.com/karpathy/nanoc…

All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges.

And more generally, any metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

the most underrated hire right now is a great product person.

when i say product person i'm def not talking about a product manager. perhaps i think there has to be somewhat of a new role. i don't have a good name for it yet but maybe something like "product thinker".. someone with an intuitive grasp of the product as it exists, where it's soft, where it sings, & how to iterate it toward something even sharper. in some sense, this person has to cohesively hold in their head where this product should be 2 years from now & work backwards from that.

i say this cuz when building was hard, engineering was the bottleneck & the status hierarchy often reflected that. building is no longer hard. which means the variance in outcomes has shifted almost entirely to judgment on what to build, how to sequence it, & how to talk about it.

& the story matters as much as the thing. internally, it organizes the team around a shared model of why. externally, it shapes the interpretive frame users bring to their first experience. you can't retrofit narrative onto a product & expect it to land, it has to be load bearing from the start.

the rarest version of this person sits at the intersection of culture & deep technology. someone genuinely bilingual. they know what's technically possible & they know which cultural currents are real vs. ephemeral. that combo is what separates products that feel inevitable from products that feel assembled.

before ppl clap back with this person has always been valuable, i know.. i am just saying now they might be the most important person in the room. their value compounds like never before.

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then:

• the human iterates on the prompt (.md)
• the AI agent iterates on the training code (.py)

The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc.

github.com/karpathy/autor…
Part code, part sci-fi, and a pinch of psychosis :)