I spend most of my time building software for internet business owners.
some oss I'm working on -
Joined April 2020
- Tweets8.1K
- Following670
- Followers334
- Likes10.1K
you starred all 7 karpathy repos and finished none
Holy sh*t Andrej Karpathy has basically put a full AI degree on GitHub and it costs nothing.
7 repos that 300,000+ developers have already starred.
Do them in order and you go from "what is a gradient?" to agents that run experiments while you sleep.
Each one builds on the last:
1. micrograd
Start here. A tiny autograd engine that shows you exactly how a network learns. Around 100 lines.
→ github.com/karpathy/microgra…
2. nn-zero-to-hero
The lecture series. Notebooks that take you from a single neuron all the way to a working GPT.
→ github.com/karpathy/nn-zero-…
3. minbpe
Before a model reads anything, text gets chopped into tokens. This is that step, written cleanly.
→ github.com/karpathy/minbpe
4. nanoGPT
Train and fine-tune your own GPT with the least code possible.
→ github.com/karpathy/nanoGPT
5. llm.c
The same training, minus the safety net. Plain C and CUDA, no PyTorch.
→ github.com/karpathy/llm.c
6. nanochat
A complete ChatGPT-style model, from tokeniser to chat UI. GPT-2 level results for pocket change compared with what it cost in 2019.
→ github.com/karpathy/nanochat
7. autoresearch
The final layer. An agent tweaks the training code, runs a short experiment, keeps what worked and loops through the night.
→ github.com/karpathy/autorese…
Most people use AI very few understand how it's built.
Work through these and you'll be the one your team turns to when the AI project lands.
same week , 1,008 mathematicians joined a group that pledges not to publish ai-written math
cloudflare's glm-5.3 on workers ai sent 155 responses with other customers' data in them . every one was a prompt over 681,280 tokens , where the kv cache pool ran out but the engine still answered with a 200
good call making it a cli any agent can use instead of one more coding agent
We launched the Gemini API CLI. A CLI for calling the Gemini API. Give it to the agent you already use (Claude Code, Codex, Antigravity, Cursor, OpenClaw, Hermes...). It's NOT a new coding agent you can use it too:
1. Generate images, video, speech, music.
2. Analyze PDFs, images, audio, video, YouTube URLs,
3. Transcribe recordings to md / json / srt,
4. Run Deep Research in the background.
Stateless. No memory, no chat mode. Your agent owns the context and the plan, gemini-api does the API call. Build for Agents, it's self-describing. `--help`, `--dry-run`...
Start here: aistudio.google.com/learn/in…
the s-1 gives the reason . they held off more than a month to tighten protection against the weights being stolen
the s-1 says the cloud plan runs over $450b , with deals that include amazon and google . both are also investors
anthropic's s-1 is public . revenue was $16.3b for january to june , more than the $4.6b it made in all of 2025 . it also spent $9b on ai infrastructure in those six months , and roughly a fifth of its revenue came from two customers
pi ended 10 of argon's 48 runs when the provider's filter refused a request . in antigravity , which retries refusals , the same model passed 26 tasks instead of 19
Gemini may have been losing to a bug, not to GPT-6 Astra.
Same model. Same benchmark. Change the harness, and Argon jumps from 40% to 54%.
Here's what happened: Pi Harness was prematurely killing runs when certain fallback or refusal paths were triggered. That affected 10 out of 48 tasks.
Google reran Argon with its Antigravity Harness.
The result?
→ Argon: 54%
→ GPT-6 Astra: 44%
→ Claude Opus 5.5: 58%
Suddenly, the model everyone thought was trailing Astra is just 4 points behind Opus.
Let that sink in.
We've spent so much time arguing about which model is smarter, but the agent framework running it can swing the result by 14 points.
If a benchmark doesn't tell you which harness was used, are you even comparing the models fairly?
#chatgpt #Google #Claude #Gemini
