Rishabh Srivastava

I genuinely think we built the best search engine for official economic data. Been working on this for 6 months. We spent ~$100k in tokens to structure economic data and make it easier to search. It's answers economic data really well. From "What has been the actual impact of AI on software engineering jobs in the last 2 years?" to "Why did egg prices increase so much more than chicken prices in the last 5 years?" Would love feedback (the more blunt the better). We have a generous free tier for the next week!

@nate_yiu Latency, yeah! I get 3 PRs with medium in the time xhigh makes 1

@rishdotblog Why not xhigh all the time? TIme? Cost? over analysis?

didn't expect this, but codex with gpt-5.5 medium has become my daily driver only situations where i use something else are: - complex backend work (use gpt-5.5 xhigh for this) - initial ideation with vague prompts (claude code w opus 4.6) - UI work (opus 4.7)

@NirantK I just run on yolo mode now, and just see the final PR 😅

@rishdotblog initial ideation, cc as a pair-programmer is pleasant

codex xhigh just unslopified a gnarly file that was a 3000 line mess. had tried everything (opus 4.6, 4.7, 5.4, 5.3-codex) to refactor this. none of those had worked without causing new regressions or race conditions 5.5 xhigh one shotted it

@mrkaran_ default for both (high for opus/sonnet, medium for gpt-5.5)

@rishdotblog Curious, what's the reasoning level you're mostly using?

Early GPT-5.5 impressions - finally an OAI model that matches Claude at tool calls Until yesterday, Opus/Sonnet were the only reliable options if you wanted to build a fast, long-running agent. GPT-5.4 was good, but thought too much, and was super slow. It also polluted the context with too many thinking tokens and so had degraded performance at long-running tasks. Gemini models are... just weird at tool calling - they often get stuck in infinite loops. GPT-5.5 costs slightly less than Opus across the tool calling loop, is just as fast, and just as good. I like its personality more - much less hedging (specially for things like financial analysis) and more to the point. It's also much more broadly useful. Codex with GPT-5.4 was pretty good at code, but Opus was just better for general tasks. GPT-5.5 feels super competent across the board. Really excited for this release. Makes the LLMs for AI-agent market competitive again!

@bygregorr @reach_vb Data viz and dashboards, mostly

@rishdotblog @reach_vb I'm not sure the image step actually helps for complex state logic though. What kind of components are you seeing the biggest lift on, pure UI or stuff with real interactivity?

for frontend agents, don't go from specs to code directly instead, to specs -> gpt-image-2 -> frontend code beats any coding agent out there. phenomenal tip from @reach_vb!

@shannholmberg @reach_vb codex has the `$imagegen` skill which can be used directly if in your own harness, use the openai api for generating an image from the spec first - then pass that image along with the spec to gpt or claude to write frontend code

@rishdotblog @reach_vb How would this work in practice? can I do this directly with codex agent?

@reach_vb jfc this works insanely well!

@reach_vb > use imagegen paired to generate what you want and then ask codex to build brilliant. asking codex to implement this rn! will report back how it goes

it's been _really_ good on the whole! main gaps with Opus: - frontend aesthetics, specially data-viz - synthesizing financial/economic data [] where I feel it's *better* than opus: - makes fewer tool calls and gets them right. really helps save on costs and latency in agentic loops - very very good at self correcting and escaping doom loops [1] take a quick look at this report with GPT-5.5 as the driver model with GPT-5.5 subagents factiq.com/share/aa0a6f13… You can look at the actual tool calls it made by clicking on the "Done in 11m" stepper + clicking through to see the trajectory. _Extremely_ solid. Couldn't be happier. But the final result (synthesizing the results of subagents) that you read in the report was off. That bit will obv improve in the next versions - but is only big gap with opus that I see in my testing so far!

The last time I dealt with model regressions this bad was in the GPT-4 era. Sonnet is borderline unusable today. Anthropic will lose so much market share if they don't secure more compute, and fast.

Something _very_ weird is up with Sonnet 4.6 rn. Tool calls are totally broken 🫠

*this* used to be impressive LLM generated code back in the day x.com/rishdotblog/st…

jfc this was less than 5 years ago slowly at first, then all at once

Welp, it's starting to happen. Under-discussed from the JOLTS report yesterday

@Yuchenj_UW Oh wow. All the best for your next thing!

I have decided to step down as CTO at Hyperbolic. Leaving a company you co-founded and poured your heart into is not easy. So many moments still feel vivid: launching our AI inference product for open-source models and seeing tens of thousands of developers sign up in a week; the week we were hit by a massive DDoS attack and the entire engineering team fought around the clock until we won; the day we launched the GPU platform and watched ARR take off. There were also hard moments. That’s the nature of building a startup. I’m grateful for all of it. What I’m most grateful for is the team. Thank you for your trust. Most startups never build something people want. I believe we did. You should be proud of yourselves. I will look forward to seeing your success. What’s next for me? I’m still figuring it out. I believe this is the most extraordinary moment in human history. We’re standing at the edge of the Singularity. AI will reshape everything, and I still feel the same excitement I felt when I first fell in love with AI. Time to start over. Time to climb another mountain. Thank you to everyone who has been part of the journey, — Yuchen

@waltertayannlee i use a refurbished m4 mac mini (bought for S$720). mostly cause I need the extra processing power and RAM for data-heavy workflows but any machine that you don't mind having to reset (because all it has is work backed up on git) is great IMO!

@rishdotblog Which sandbox do you recommend btw?

Sigh does everyone saying "it will be expensive to maintain AI generated code" not know about --dangerously-skip-permissions Give Claude Code / Codex a sandbox, let it poll for updates/canges, auto push and review PRs, set up CI/CD. Automated, better-than-human maintenance 🤷🏽‍♂️