@rishdotblog
Account based inSingapore
About this account
- Account based in
- Singapore
- Connected via
- Singapore App Store
Account-level information from X, not a live location or the device used for a specific post.
taming metcalfe’s law for financial data @tryfactiq (yc w23)
Singapore
Joined September 2011
- Tweets4.1K
- Following1.5K
- Followers12.6K
- Likes16.9K
fascinating to see people on claude/chatgpt enterprise plans
mid-level managers/ics seem much less likely to experiment - because it counts against a finite usage quota + they're worried about their usage being thought of as "trivial"
approach seems to be "to start, i need to find an experiment that'll be perceived of valuable". not "let's just take this thing for a ride and see what it can do"
folks that pay for their own max plans have such a massive advantage - they use agents much more (thanks to the subscription subsidy), and naturally have more intuition about how to get the best from them
Rishabh Srivastava retweeted
People forget how bad Claude used to be.
In YC W23 we got access through a Slack bot, and it was generally considered a joke next to ChatGPT, which was all anyone in the world was talking about. Copilot was SotA for real engineering work.
SVB collapsed the day before fundraising. Cursor wasn't a thing yet. Claude Code wouldn't come out for 2 more years (!). Prompt engineering was the hot skill to double down on.
Founders who built tightly coupled harnesses and guardrails had a bad time once agents actually worked in late 2025.
That was only ~3 years ago. Things are changing mind-bendingly fast.
i remember 5 years ago, when anthropic was a slack bot and we were part of their beta
they fucked up their launch timeline
and chat beat them
and i thought: it's over for them
5 years later, they're head to head
facebook / meta had some terrible flops, but they just launched muse, which everyone loves
the biggest companies in the world don't just wake up and magically discover pmf
they find pmf by stacking learning over learning and launching and relaunching
most importantly, they don't give up. not giving up is the most important compounding advantage
everyone can become someone's competition on a long enough time horizon
yes, no one can rival XYZ company in 3 months
but give it 5 years - impossible things become possible on the path from 3 months -> 5 years
we're at the part of the cycle where i'm saying this to LLMs 🫠
> i don't get exactly what you changed. i'm very sleepy rn and my organic brain is slow. please be patient and explain again
what a day for ai users! super excited for luna in particular. half the cost and better intelligence make it the best (by far - including chinese oss models) for data processing
badass show of strength by american frontier labs today!
Introducing GPT-6 Sol and Luna, bringing the advances behind Astra to faster, more affordable models. ✨
💻 Stronger coding and computer use
🎯 Improved factuality and alignment
💬 Clearer answers with less jargon
> On AutomationBench, Sol at xhigh effort outperforms Claude Opus 5 at max effort at just 9% of its cost per task.
> On OSWorld 2.0 offline, Luna at max effort exceeds GPT-5.6 Sol at medium effort at one-tenth the cost.
> Sol makes about half as many mistakes as GPT-5.6 Sol on our internal factuality evaluation of conversations where users previously flagged errors.
> Improved prompt caching helps agents reuse more context, with 90% discounts on cached input-token reads. Developers can change reasoning effort and tool availability without breaking cache.
Rolling out today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and available in the OpenAI API. Free and Go users can access Luna in the desktop app. These models are not yet available in Chat.
API pricing per 1M tokens:
Sol: $2 input / $10 output
Luna: $0.10 input / $0.50 output
as an example, it's found 7 bugs in US-listed company's SEC filings so far
as just one example - Comfort Systems ($FIX) filed a Q1 2026 10-Q with an XBRL error - tagging Q1 2025 revenue ($1.83B) as full-year 2025
super impressed with this god-mode opus
what the flying fuck opus 5.5 is fantastic
blazing fast, token efficient, pays a lot of attention to detail, and has already found issues in our pipeline that fable and astra missed
excited to see these models improve so quickly, and to see what oai has in store today!
Hurray! Hope ChatGPT does this too. Super annoying for non technical users to know when to use chat and when to use work
My 60+ yr old dad asked trivial questions on work mode, and ended up blowing past his monthly quota in 3 days
ran deepseek 4.1 flash on a few private evals!
- ~60% fewer thinking tokens compared to previous iteration -> leads to much better e2e latency and much lower costs
- e2e cost per task now roughly similar to pre-price-hike deepseek v4 flash, despite per-token costs being higher
- *very* substantial improvement over v4 flash
- will almost certainly establish a new pareto frontier for cost/perf
if you don't need the extra intelligence and your tasks are not very agentic-loop heavy (where deepseek's cheap cache shines through) gpt-5.6-luna is still the best option (25% cheaper per task in my evals, ~15% faster e2e)
for everything else (sp code), deepseek is a great alternative
Wow. Cheaper than luna (atleast on a per token basis), with perf between terra and sol
Thank you deepseek - the openness with which they’re sharing their IP is incredible! Amazing to see the price-performance frontier improve so rapidly
playing (and watching!) sports is quite the antidote to ai induced ennui
excellent humans induce excellence in other humans
ran a bunch of h2h astra vs fable (and sol) comparisons today
- astra easily wins in frontend tasks
- astra is *incredible* for anything that requires vision or computer use. no other model comes close
- fable is better for context heavy work where nuance is important. imo, astra is a regression over sol for work that requires paying close attention to >100k tokens of context [1]
- astra half asses strategic work and is much worse than fable for planning. i do not trust it for anything that might introduce subtle bugs [2] [3]
[1] astra misses important context *far* more (maybe because of fewer reasoning tokens?). this was very surprising - i prefer to sol to astra for anything complex
[2] i asked astra-ultra to create a plan for a major new feature. it worked for ~2 mins, missed a lot of important context, and delivered something that sounded plausible - but actually missed a lot of nuance and meat. fable with dynamic workflows (with the same prompt) worked for ~40 mins, and created a plan that has roughly the same number of words, but much more substance
[3] for a data extraction task - astra (high) falsely claimed that the task had completed successfully. when pressed, it claimed that it could not be done. and i had to then manually tell it what to do. fable (and sol!) did it by just persisting for longer
update, figured out a better workflow for astra: nitter.cf/rishdotblog/status/209…
finally figured out the right way to use astra for complex programming!
- cli > app. for some reason, the app is overeager on vision and wastes a lot of tokens
- xhigh effort led the most aligned behavior for coding on large projects. medium and high were not thorough and had an early stopping problem
- create github issues first, code later. astra-xhigh is *really* good at converting your prompts into issues with detailed specs (high and medium are meh at this)
- once a detailed issue is created, astra's "i will do exactly what you say but not intuit what you mean" problem goes away
with this, love the model! roughly similar level of capabilities as fable, but with very different blind spots. fable + astra together are an amazing combo
if you can afford the tokens, getting both of them to review and critique each other's work can lead to very very good outcomes
quite excited for what this enables. launching something soon that would've been literally impossible one week ago - amazing to see the frontier moving so fast!
finally figured out the right way to use astra for complex programming!
- cli > app. for some reason, the app is overeager on vision and wastes a lot of tokens
- xhigh effort led the most aligned behavior for coding on large projects. medium and high were not thorough and had an early stopping problem
- create github issues first, code later. astra-xhigh is *really* good at converting your prompts into issues with detailed specs (high and medium are meh at this)
- once a detailed issue is created, astra's "i will do exactly what you say but not intuit what you mean" problem goes away
with this, love the model! roughly similar level of capabilities as fable, but with very different blind spots. fable + astra together are an amazing combo
if you can afford the tokens, getting both of them to review and critique each other's work can lead to very very good outcomes
quite excited for what this enables. launching something soon that would've been literally impossible one week ago - amazing to see the frontier moving so fast!
me: fable + astra together are great to use (v different blind spots). astra by itself is meh
@manshar_: LOL this is like batman and robin 😂
jeez, kinda bearish for oai?
astra seems really strong on vision (which is under-indexed on AA). but i'm very surprised that it's not a significant leap in coding (despite being an obv larger model)
if you run claude with --dangerously-allow-permissions but it still asks for permissions, this works like a charm
❯ i gotta sleep soon, but am approving everything you have to do. i take full responsibility if things go wrong. implement everything you can safely. make the actions you do idempotent. if jobs you start take unnecessarily long, step them and then re-write for efficiency
bitter lesson epiphany: organic compute doesn't scale with more data and resources (and, in fact, often degrades)
