Fernand Pajot
1.4K posts
@rfleury @typesfast No. The model learns a *distribution*, this is a common mistake people make. You can get it to prefer whatever is in the distribution.
English
With all these AI coding improvements why isn't the software I use everyday getting better?
English
@mitchellh Everyone is liking this but very few people will do it :(
English
Fork your dependencies, trim them to only your use case, never update unless it breaks for your users. I’ve been vocal about this for 10+ years. I’ve always said that updating is way riskier than latent bugs (which can be tracked and CVEs monitored).
If you are updating a dependency, it’s on you to analyze every single commit in the full transitive set of dependencies. If you dont see anything compelling, dont update!
I remember at HashiCorp once in awhile an engineer would try to update a dep or replace a DIY lib with an external one and id always ask “show me the commit we need.” Dont update for the sake of it.
Feeling pretty swell about this mentality with all the supply chain attacks happening.
English
ironically think it’ll be a sad time for ai researchers this year. they are first in the hotpath of RSI and probably the market for them will shrink or at least their pricing power will be reduced as this generation of models commoditizes the skills that made them rare
English
I remembered that I HATE triangle LODs, so I started toying up with point splatting ideas. The original Gaussian Splatting paper used super bloated 248B splats. Best modern versions seem to be 16B-48B per splat. Still seems quite big as splats are often 1 pixel in screen...
English
@paulg Very misleading graph. If you look more closely the regime change happened first during the 08 crisis. Smartphone penetration was still very low.
English
If Steve Jobs were still alive, he would have the moral authority to face and maybe even to solve this problem. But I doubt anyone in the phone business now does.
English
@KIRI_Engine_App Have you done experiments with TriangleSplatting 2?
English
3D Gaussian Splatting Mesh 3.0, more results coming in.
The main improvements: smoother large surfaces, better structure retention on thin geometry, and less broken mesh overall. Side-by-side with 2.0 in the video below, thin details is where it really shows.
Full showcase: kiri-innovation.github.io/3DGStoMesh3/
This is what the algorithm change actually looks like in practice.
English
I feel like most 3DGS demos I’ve seen are of larger environments and look great until you zoom in to anything, upon which everything dissolves into a radioactive cloud.
This on the other hand is the most amazing 3DGS demo I’ve ever seen.
Dany Bittel@DanyBittel
A Raspberry. 90 stacks, 68 photos each. 2.37M splats. #3dgs
English
@eliebakouch Yeah saw those, mostly interpolations/riffs off existing "big" ideas, though decent taste.
English
@fernandp you can see some of the ideas from novelty track here! but yeah they didn't beat the record indeed
claude: github.com/PrimeIntellect…
codex: github.com/PrimeIntellect…
English
we let opus 4.7 and gpt 5.5 run on the nanogpt optimizer speedrun: ~10k runs, 14k H200 hours, 23.9B tokens. opus hits 2930, codex 2950, both beating the human baseline of 2990. we cover claude autonomy failures, codex high compute usage, and much more
primeintellect.ai/auto-nanogpt
Prime Intellect@PrimeIntellect
Automating AI research is the next major step in AI We let Claude Code (Opus 4.7) and Codex (GPT 5.5) run autonomously on the nanoGPT speedrun optimizer track using our idle compute. ~10k runs, ~14k H200 hours Opus now holds the record at 2930 steps vs the 2990 human baseline
English
It's now even easier to keep your agents running productively 24/7
OpenAI@OpenAI
You've been asking for this one... Now in preview: Codex in the ChatGPT mobile app. Start new work, review outputs, steer execution, and approve next steps, all from the ChatGPT mobile app. Codex will keep running on your laptop, Mac mini, or devbox.
English
@Love2Code Codex has been taking screenshots of my C++ apps by itself for me on Windows. It's been hit an miss in terms of getting it unblocked though.
English
I'm finding that asking LLMs to debug 3D graphics problems can be an experience in frustration. They can't see the program running so you have to describe the issue to them. They speculate as to the cause of the issue and are often confidently wrong.
English
@leooeld I didn't know you could get so high as 84% on Hopper, that's impressive.
English
I bet they used BF16-throughput as the denominator when training in FP8 or something. By that algebra, I can get you 150% MFU in no time😅. For reference, as far as I know the SOTA Hopper GEMM kernel is ~84% utilization.
arxiv.org/abs/2605.05331
English
@yacineMTB Heart rate monitor chips cost around the same amount as an IMU chip
English
@MikeIsaac This is exactly what my toddler has for snacks and lunch haha, minus coffee
English
lunch for court tmrw: turkey sando, adult gogurt, fruit leather, coffee
English
@wieslawsoltes @SheriefFYI Why is egui better for user facing stuff than imgui?
English
@SheriefFYI Try github.com/emilk/egui , check #why-immediate-mode" target="_blank" rel="nofollow noopener">github.com/emilk/egui#why…
English
are there UI frameworks that are immediate mode and user presentable? imgui is great for debug UIs but shipping a product with it leaves a lot missing. Valve has a great D3D11-rendered UI for things like SteamVR, does something open source like that exist?
English
Unless you are early at a startup, you should be job hopping as much as possible
Every time I job hopped I've doubled my income. If I didn't switch jobs every time someone wanted to hire me, I would be much much poorer today
Aidan Clark@_aidan_clark_
One of the things I like the least about SV is how little time people stay at jobs. Important systems are complex and understanding them requires watching impacts over long time periods. I’ve been at OpenAI almost four years and I still feel like I don’t understand much of it!
English