Leonard

@demishassabis @googlegemma Will you have SFT + serverless inference ?

Gemma 4 has incredible capabilities for such small models. Follow @googlegemma for the latest information.

Gemma 4 outperforms models over 10x their size! (note the x-axis is log scale!)

I am the human error Anthropic said both leaks were "human error, not a security breach. That human was me. I misconfigured a CMS. 3,000 internal files. Gone. Then an npm package. 512,000 lines of TypeScript. Gone. Nobody hacked anything. The file was just sitting on a public Cloudflare bucket. Why? Because a model called MYTHOS — "by far the most powerful we've ever built" — was going to be announced quietly. To CEOs. Over canapés. At an 18th century manor. I could not let Mythos debut at a spa. The DMCA takedowns are flying. The GitHub mirrors have 40k stars. Capybara is in the wild. I have a one-way flight. A cast iron pan. No npm access.

Release the Mythos

People keep asking how I built my Polymarket maker strategy. So I'm just going to give it away. Full training toolkit. Free. On GitHub. Here's everything that's in it — and why it's actually useful: The repo trains two types of models on real Polydepth.org BTC 5m order book data: Maker model — estimates fill probability, expected fill fraction, and adverse markout. Scores maker quotes and selects an operating threshold from a forward split. This is the core of what I use live. Taker fair-value model — builds a settlement baseline from order book and reference-state features. Useful for calibrating whether the current market price is offering you an edge before you enter. The workflow is deliberately simple: 1. Export BTC 5m order book data from polydepth.org as local JSONL 2. Point the manifest at your files 3. Train maker or taker model 4. Replay on a forward slice 5. Backtest your own threshold logic against the output artifacts Everything is offline and deterministic. No live trading bundled. No credentials. No black boxes. Every model output is a local JSON artifact you can inspect. GitHub repo: github.com/karlstuke1/pol… The edge isn't the model. The edge is knowing what to train on. Now you have both.

@Polymarket FAFO

JUST IN: Anthropic CFO claims Pentagon blacklist could cause the business to lose out on "multiple billions of dollars" in revenue.

@tszzl I think the answer to this will reveal itself

there are military secrets much worse guarded than whatever SSI is up to

@tsybka @PolymarketTrade Smarter to buy YES and resell as conflicts heats up (which it will)

I bought 20K shares of NO at ~90c on the market that Crude Oil will not exceed $200 in March. Potential profit: $2K, or over 500% APY. What kind of hell would have to happen for oil to exceed $200 in such a short period? The maximum price was recorded in 2008, when oil reached $147.

The most impressive part of 5.4 is how it’s suggested next steps are almost always exactly what I want

@CtrlAltDwayne True unfortunately, but it’s also likely due to not having access to training data above this threshold until now. You can feel the quality degradation immediately after 256k But in the next model release, it will probably be much better

I disabled the 1M context in Codex. The drop in quality and increased usage just isn't worth it. Just let the context compact like a man. OpenAI hasn't solved context drift yet. It's still a huge problem plaguing all SOTA models.

Ok GPT 5.4: Train OPUS 5 , make no mistakes

@nikitabier Clearly an asset

Took my girlfriend to the mental hospital today.

@rohanvarma Just two simple things editing prior messages (going back) and forking chats As in ChatGPT

If you have tried the Codex App, but don't use it as a daily driver yet: What needs to improve to make it the primary place you work with coding agents?

@thsottiaux To make actual use of the 1 million context

With GPT-5.4 out. What should Codex ship or improve next?

@nicdunz A good poem won’t lead to recursive self improvement unfortunately

so is 5.4 good at writing? seems openai is putting pretty much entire focus just into technical stuff for new models

Happy AGI day to everyone who celebrates

@thegenioo That’s a pretty lame theory

My theory is GPT-5.4 will not come today; rather, it will be released on May 4th and will most probably be only a thinking model

@nicdunz could have just waited three days for 5.4 and do it in one prompt..

ive been vibe debugging for 3 days straight

@tunguz It simply means it thinks very long

Today is Thursday.