DataThoughts

@leerob It’s great seeing the Composer 2.5 progress, but benchmark with Claude 4.7 Max at #1 is a bit suspect based on my own experience

Where could we improve Composer 2.5? We're working on the next model and would love your feedback. Lots of work to do (our CursorBench evals below) in the coming weeks!

@thsottiaux seeing major 5.5 degradation in codex output quality today - completely ignoring prompts, not answering what I’m asking, early exits, etc. It’s been cooking all week so this regression was quite noticeable. Love the app/model - just frustrating hitting a wall

@iuditg Claude Code and Codex are harnesses - there are plenty of great OSS harnesses.

Unpopular opinion: Open source will never be as good as Claude Code and Codex.

@ohryansbelt @karpathy If Opus 4.7 wasn’t a huge regression I’d agree, but I suppose they have mythos.

@karpathy this is like KD joining the 72-9 warriors

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

@dansk_holger @financedystop If you don’t touch it, it will grow to almost $1.2M after 10 years, $2.5M after 20. She’s only 20, if she held it until retirement (65), that would turn into $17.5M. Big if though.

@financedystop It's not $1 million net. After taxes it is more like $550,000 At 8% that's roughly $44,000/year

$1 million invested at an 8% annual return could generate roughly $80,000 a year on average while still growing long term. That’s why most financially literate people would take the lump sum

@MurrayHillGuy1 1, 2, 7, 14, 15

If you're a millennial it's time to pick your midlife crisis: 1. Quitting alcohol 2. Running 10 miles before work 3. Divorce 4. Panic baby at 35 with wife you hate 5. Pickleball 6. ADHD diagnosis 7. Dressing like you did in 2004 8. Blacking out every weekend like you’re 21 9. Weekly hinge dates 10. Ice baths and saunas 11. Board games and craft beer in the suburbs 12. Getting into tattoos 13. Quitting your job to explore your “passions” 14. Plants and the environment 15. Traveling

@VictorTaelin I think so - they don’t _have_ to release to retain market share, they’re printing money. They release when there is meaningful improvement, which should give you hope for their next one.

should I have any hope Google launches something comparable to Mythos 🥲

@_TipsTricks You just need to crumple it into a ball and it’ll fit anything

I learned this from Grandma

@kr0der I’d probably give it to our best engineer

if Anthropic hypothetically released Mythos on a $500 or $1000 subscription, would you subscribe?

@zuess05 A real software engineer has scar tissue. They can knock out a quick POC over the weekend too, but they code defensively because they remember all the ways it broke in the past, or the nights they were pinged at 3am to fix an issue in prod. That 19 y/o hasn’t had to firefight.

Serious question. For 20 years, a "Software Engineer" was someone who spent thousands of hours mastering complex syntax, logic, and architecture. Now, a 19-year-old can vibe-code a production-ready SaaS in a weekend using plain English and a $20 Claude subscription. What does the title "Software Engineer" even mean right now?

@HazelAppleyard automatic transfer to a brokerage or create a shell company and charge myself $10M for “services”

How would you manage it?

@robinebers I had to cancel Claude, it does what it wants and likes to delete shit it shouldn’t - all in on codex for now until Opus 4.8 then I’ll re-eval

goodbye codex unsubscribed from the $200/mo plan for now → love the team → love the app → love the intelligence of the model but my business isn't just about backend code it's a lot of landing pages. a lot of emails. a lot of content, intros, scripts, tweets, and instagram carousels. none of which codex is good at

@Rothmus The only states I want to live in

“10 worst states to live in” and they’re some of the best states to live in if you have a family

@krzyzanowskim How are you implementing the spec? Is it iterative? What guardrails/hooks do you have?

"skill issue" - a whole day of writing PRD/specification - half day grilling the PRD/specification - 16h /goal implementing PRD it doesn't work. IT DOESN'T WORK. it doesn't work at all, but also it doesn't work as specified. why broken? "So implementation drift" I'm done with this shit!

@TheVibeShift @yacineMTB I have a super strict clippy config and feed the goal a well define spec with an iterative planning, implementation, and review workflow. It probably only codes in 20 minute increments before going into compilation, unit/integration/e2e/performance tests, replanning

@wishful_data @yacineMTB I can only imagine the garbage you guys are producing. Even 20 minutes is too long to leave an AI unattended.

Never use a codex prompt without /goal by the way

@AishwaryaDevv Add more guardrails: - super strict linting - project structure framework - docs framework - cve/dependency audits - live e2e checks Retroactively apply them via a painful /goal run and make it automatic going forward

Am I the only one getting vibe coding fatigue? Building landing pages in 30 seconds was fun, but maintaining a complex codebase where half the logic was “vibed” into existence is an absolute headache. Feels like we traded 1 hour of typing for 5 hours of architectural debugging later. I’ve started manually writing core logic again so I actually know where the technical debt is hiding. Is anyone successfully managing large production projects with AI agents, or are we all just building disposable software?

@sporadica Anything you say or do can be used against you in the court of law

Maybe it’s the Gen-Z in me, but i fully don’t care about privacy. I am post-privacy. I am giving OpenAI access to all of my finances, all of my health data, everything, I don’t care anymore

@mcuban As long as the prerequisite for this tax is that the government must balance their budget

We should federally tax Tokens at the Provider level. Not a lot. Less than 50c per million tokens. It will accomplish 4 things (at least ) 1. It will push the big AI players to optimize tokenization, caching , routing and localization Which will 2. Reduce energy usage. Saving them in energy costs more than what they paid in tax and reducing strain created by the growth in energy consumption Which will 3. Generate maybe 10 billion dollars a year to start, but over the next ten years could grow 30x to 100x Which will 4. Create a source of funding to pay down the federal debt or deploy, in response to the things AI brings that we don’t expect or don’t like At some point the models will pass it on to customers. Of course. That’s ok. Customers will have the ability to choose between providers. Or to do everything using open source models locally. Thoughts ?

@ibuildthecloud I hit the $200 weekly limit today, /goal was running for 6 days

I'm using the codex $200/mo plan. I never hit limits, I'm going to switch to $20/mo, see how long it takes to hit the limits, then try the $100/mo plan. I think I might really only need the $100/mo plan.