Austin Ray

The early days with @bcherny and Cat were so much fun. I remember joining our first call together and kicking things off laughing out the words "Oh my god, you guys, how crazy is this, right?". We shared what we'd been hacking on, went over some bugs I'd reported on Slack, shared tips & tricks. It was a surreal experience - I figured Claude Code was likely a black swan event, and, somehow, I was talking to its creators, not even a week after it was released. I honestly couldn't believe my luck. I snapped out of it at some point and remembered what I was there to do: "Everyone at Ramp needs to be using this. I'm going to do everything I can to make that happen. Can I count on your support?" It turns out: yes, definitely As Boris likes to say, the cultures at @tryramp and @AnthropicAI are really similar. Values, pacing, culture. Our engineers started nerding out together pretty much immediately, and the partnership blossomed from there. By the end of May 2025, most Ramp engineers were using Claude Code. 16 months after that first call, I still feel lucky to be a part of this. To me, what's going on with AI right now is the most exciting thing that has ever happened, and I feel grateful every day (1) to be alive right now (2) to be able to work at a place like Ramp (3) to be able to work closely with the folks on the Claude Code team (shout out to Adam, Daisy, Karl, Igor, Moez, Adeline!). No doubt the next 16 months will be more bewildering than the last. I can't wait :)

AI lab revenues are going to be really volatile. Being agile enough to surf the waves of the cost/intelligence Pareto frontier will be important for staying competitive as a business. As the market gets more efficient, the lab with the best models at the best prices will get a higher and higher share of AI spend. And then experience worse crashes in the other direction. Switching between providers is easy, and the better models help you do it faster and with less human effort and disruption.

I have been thinking about the famous chart showing how experts keep projecting linear growth in solar installations, year after year, and always get it wrong when growth is still exponential. I think the same thing is happening with the discourse on product strategy around AI.

Yeah we’ve caught glimpses of this type of behavior at Ramp. Nothing bad so far, but the reports we’re seeing on X (eg the rm -rf ~/ guy) don’t seem that outlandish. Soldozer is a bit of a wild card.

@sama @steipete 🎵Tale as old as time🎵🌹

there are a lot of benchmarks that suggest 5.6 sol is the best model in the world right now, but the most reliable way to tell is that elon is obsessed with me again

Supposedly an internal memo by the GLM CEO. The entire thing is a great read.

With 5.6 Sol, a lot of people are still prompting the model exactly as they did 5.5 It's important to note that 5.6 Sol is a lot more tenacious and thorough than previously models. Check out the guide I wrote here to get better outcomes learn.chatgpt.com/docs/prompting

The prompt engineering here is super impressive! Such a great example of agent prompting:

Yessss finally! Radical improvements to human health incoming. Wearables market gonna pop - much deserved

Yikes - the authors seem to misunderstand where we’re at with coding agents. 1. They say that 2028 (!) will be when AI is better than the median software engineer.. 2. They have it down as 2029 (!) being when we automate coding.. 🤨 1 was early/mid 2026 2 is mid/late 2026

Epilogue got me like 🫪

@DKokotajlo Wild. What a completely, utterly wild time to be alive. Thanks for putting this together

In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power. In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.

Congratulations to @DKokotajlo and team on releasing Plan A (ai-2040.com) , a followup to AI 2027. My attempt at a gentle introduction to the ideas involved is at astralcodexten.com/p/introducing-…

The replays of Sol on ARC-AGI-3 are cool to watch. Few examples of Sol max: 1. arcprize.org/replay/f018164… 2. arcprize.org/replay/2474010… 3. arcprize.org/replay/6693f6a… The aggregate scoring is sort of weird. On specific eval games though, it's interesting to see how big of a jump there can be from one effort/reasoning level to the next. Inference-time "grokking" in action, I think.

TLDR: Luna + Sol dominate cost vs intelligence pareto frontier (i.e. they currently sit atop the valuemaxxing throne). ARC-AGI-3 says Sol (xhigh) is first model they've seen show sparks of genuine fluid intelligence comparable to that of a human (possessing adaptable, creative problem-solving abilities) -- caveat: Fable not tested by ARC-AGI-3. AA benchmarks show Sol as cheaper and slightly better at coding than Fable, but slightly dumber in general. 1. artificialanalysis.ai/articles/gpt-5… 2. arcprize.org/results/openai… 3. x.com/arcprize/statu… Artificial Analysis quotes: > GPT-5.6 Sol (max) in Codex leads every evaluation in the Artificial Analysis Coding Agent Index > GPT-5.6 Sol (max) scores 1 point below Claude Fable 5 (max) in the Artificial Analysis Intelligence Index at 59 points, at approximately one third of the cost. > for any Terra effort level, there is a Luna or Sol effort level that is more intelligent at no extra cost, or equally intelligent at lower cost. ARC-AGI-3 quotes: > GPT-5.6 is the first model to show material progress on ARC-AGI-3. It is the first model to win an ARC-AGI-3 public game [in] 152 actions compared to 208 human baseline > GPT-5.6 Sol is the standout model of the GPT-5.6 family. > Sol is able to read an unfamiliar scene correctly and in the game's own vocabulary. It treats a failed hypothesis as a reason to re-plan rather than thrash. Most agent failures are upstream of the code they write or the action they take. Sol is able to perform on ARC-AGI not because it executes better, but because it correctly orients itself in a new environment first. > Sol’s distinguishing capability is scene comprehension. It almost always figures out what the core game mechanics actually are (unlike other models). When it loses, it loses downstream (planning/execution), never at the perception stage

Wow. Look at that jump from Sol@high (2.1%) to Sol@xhigh (7.0%). Looks like Sol@xhigh gained emergent capabilities in some of the games, presumably by crossing some sort of inference-time compute threshold required to "grok" the game. Fascinating.

Hint for all AI Labs as they branch out from work for programming to general knowledge work: non-coders are not just dumber coders Taking away a bunch of options from your coding app does not make it better for knowledge work. We need more types of control & visibility, not less

You can now vibe code a language model. From a single prompt, GPT‑5.6 built the entire training pipeline and trained a model from scratch on my iMessage history. Locally on my Mac. It now generates replies in my writing style.

Welp it's ai christmas again