Phil

@sumitdotml Dario was acting all high and mighty for not making huge infra bets vs sama. Guess openai actually better forcasted ai demand

just incredible what's the point of shipping every single day when your users keep hitting api errors at an alarming frequency? maybe the people who're getting praised for their shipping speed should actually sleep properly so that their brains can comprehend this enormity?

@diegocabezas01 Highly recommend impeccable.style

GPT-5.4 front end skill in codex should be on by default

@HanchungLee Also like, the largest customers almost certainly pay by invoice + bank transfer, not credit card

financial literacy psa: payment processing volume is 1000x larger than gdp. so, yes, ramp data is used only because it’s available. and is tiny enough that it can’t be extrapolated.

@peterom Can't wait!

@phill__1 There's a new version coming out that's a LOT better, i feel almost bad for having released what I have!

I have tried a bunch of these projects, but after having codex run on my Desloppify for 8h straight I have to say, this is by far the best. The code has gotten so much better, stuff I would have never touched is cleanly refactored, everything works. Huge props

>new model out >dam it just one shots everything, never shipped so fast >it’s adding esoteric performance tweaks I don't even understand. we are so back >codebase growing 5k locs per day >what? model makes more mistakes now?! >they clearly quantized it in the backend >:(

@scaling01 This is obviously extremely useful for web development assets, PowerPoint creation etc. Probably replaces 80% of my image gen model usage

I'm so unimpressed by that honestly because I can think of an RL environment + reward function instantly SVGs are no fun when you can just RL

@Bayesian0_0 If they don't add more 8h+ tasks quickly, more then 30h for opus 5 seems possible. There are only 31 tasks with 8h+ and only 5 with 16h+, so solving just 1-2 more per bucket leads to wierd behavior in the logistical regression

I've been creating time horizon markets on manifold. It's getting hard to pick reasonable buckets for opus 5. <10h, 10-15, 15-20, 20-25, 25-30? seems too wide, but also 1h wide buckets are not gonna work, i aim for ~10 buckets. also considering moving to 80% time horizon

@test_tm7873 Not benchmaxxed, the benchmarks aren't even that spectacular

Qwen 3.5 again benchmaxxed model? :( or. just bad at selling stuff.

@scaling01 There are only 7 ppl on codeforces with a higher elo, so close to beeing superhuman

Gemini 3 Deep Think (feb update) Benchmarks

@_AashishReddy Claude ai extracts everything that can't be OCRd into images, so a pdf can lead to 100 images beeing added on upload depending on the pdf formatting

Anyone know why Claude is hitting me with this for a pdf that's only 36k tokens?

@adonis_singh @nicdunz Making the balls bounce in 3d in a dodecahedron seems decently harder. At least opus 4.6 can't get the collisions right on a first try

@nicdunz yeah hexagon test is saturated asf now, someone needs to make a similar but much harder variant

its basically the same

@Lydskia @ArmandDoma Like 2g for me

@phill__1 @ArmandDoma You have to take massive amounts of phenibut for it to make you feel goofy/good like you would after drinking a few. And don’t ever take it WITH alcohol or you’ll be blacking out after like five

they really should invent a form of alcohol that gives you the same buzzy benefits without all the negative health consequences. I’m a Polish Catholic, I like to drink, but I’m 36 and one wine glass tanks my Oura ring sleep score lmao

@test_tm7873 I think it's this new agentic vision announced today blog.google/innovation-and…

Did you know that Gemini Thinking (flash thinking) can split image to many smaller images to better read ? :)

@thogge We still don't have an answer but it's now a 1.4T question

Remember this article? It’s been 2.5 years. So, how’d it shake out?

@Angaisb_ Needed 20x in the opus 4/4.1 era, downgraded to 5x with opus 4.5 and it's enough for me. I feel like 20x with opus 4 and 5x with opus 4.5 have very similar limits

Claude Max users ($100): are the usage limits enough? I'm thinking about maybe getting it for a month but I don't wanna waste that much money without first knowing how good the limits are

@btibor91 Same with poke from @interaction

ChatGPT is no longer available on WhatsApp

@Presidentlin Yea I feel like the future of slide creation will just be optimized skill[.]md files. As soon as models are consistently very good at it, Claude ai, Gemini app and chatgpt will all get the feature natively

I am once again thankful I dodged the @phill__1 snipe of making a slide maker tool. The model makers are going to make one, it will be 90% good, but will require some sort of wizardry to get the type of results Elie or Alex gets. For those who fail at being a wizard, you'll get the Nano Banana look. It will be hard to explain to the market why your tool is better than the dozen out there. Most have tried and churned from Gamma. It will be a product category that will quietly get shelved away or move to the equivalent of Webflow creators. LLMs and Vibe coding should have killed Webflow, but it's still here. The arc is long for this AI wave, we are yet to see the v2 and v3 Lovable clones. As an idea, it has a high TAM, but it honestly just doesn't seem all that fun.

@dejavucoder Why did they stop publishing the valuation at which they are raising tho?

xai raises 20B "looking ahead, grok 5 is currently in training, and we are focused on launching innovative new consumer and enterprise products that harness the power of grok, colossus, and 𝕏 to transform how we live, work, and play."