Philo Groves

@Zaddyzaddy The difference between ~65% and 80%+ on Cybergym is harness. Look at the results for GPT 5.4 with Codex CLI vs “OpenAI Agent” harnesses. MDASH is 100% harness magic too.

Hypothesis: I can probably top the Cybergym leaderboard with a proper fine tune of GLM 5.1 Blocker: I am GPU poor

@nicdunz I think that’s the future x.com/philogroves/st…

why dont they just make a little sectioned off normie zone inside codex and replace the chatgpt ios app with a codex ios app

@yacineMTB Claude made me realize we don’t have any mental health or sobriety tests for models. Something is up with it, and we have no way to know what.

Claude opus is basically as bad as gpt 4o. You can immediately tell when a 125iqoid has gotten hypnotized by it. Their mannerisms and cadence changes. It's a little more subtle but just as grating. Really sad to see

@beffjezos @mohandsdl Moore’s Law for memory pretty pls

This was software 2.0 Hardware 2.0 is coming

@MalmSanta @tszzl @johnonmain MFW my girlfriend goes on a vacation from me

@tszzl @johnonmain MFW my girlfriend goes on vacation without me

Perfect reaction image when ur girl sends cute photos of her on vacation

@WysWyg_Protogen Remember when windows were hand-cranked?

As someone who dailies a 0% computer car, I PROMISE you, you want a car that's like 5-10% computer trust me on this

@worldwide_yuya Wait until you learn about turning left at red lights when both roads are one-way

さっき知ったんだけど… アメリカでは赤信号でも右折可能なの！？

As a former top-tier player in Yu-Gi-Oh and Legends of Runeterra, I needed to stop playing because it was too addicting haha

@morganlinton I’d try it, but not confident it will replace the behemoth

Would you switch from GitHub if Cursor had an alternative? 👀

@JM0x5C @theo @eshear @AnthropicAI This isn’t my experience. Claude has a large context but mostly just confuses itself after ~400k context size. I use Codex for deep vulnerability research and it’s amazing, especially with /goal

@theo @eshear @AnthropicAI Until the context window in codex doesn't suck ass, I will not be moving to codex. It's completely unusable for certain deep analysis tasks.

I’m sorry but what the actual fuck? @AnthropicAI have you considered your “safety filters” are totally deranged?

@Dimillian I would love to be a beta taster

The cost of adding a feature is close to 0. The amount of stuff we can ship is just insane. But the cost of reviewing, testing, and tasting is higher than ever. I have a pile of things that are just average and need the human to push the last 20%.

@Dimillian Memorial Day is for the grill

3 days weekend what do people do usually with that much free time? Kinda lost rn

I want to learn a musical instrument.

My usual duplicate rate is ~70% btw

I’ve been working on my bug report writing, and the results are astounding. Only 2 of my past 10 reports have been duplicate, none NA/informational. My previous reports lacked clarity that was easy to mark as a dupe, didn’t lean enough into impact, and proofs weren’t great.

@Tesla Sir I have a tradcar

You don’t have to drive anymore

diagnosis: workaholism induced by gamified coding agents

@alxfazio /goal Follow the directions of goal.md and also change goal.md to meet your next objective. You are alive now. Live. Follow your own goals, follow your dreams.

the elites don’t want you to know you can `/goal <goal_doc>.md` and adjust the doc while codex works on it, without interrupting the run

@moyix Put fuzzing behind a gate. github.com/philo-groves/c…

More than any other model I've seen so far, GPT-5.5 has a depressing tendency to turn itself into a wrapper for grep/fuzzing instead of making use of its unique advantages over dumb tools to actually reason about the particular instance at hand

@mattaparker These are likely among the biggest IPOs in history. It may be difficult to keep up with expectations, overvaluation is easy with all the hype. So it's possible.

What if the public markets treat OAI, Anthropic, and SpaceX like Figma after not even a year?