David Picard reposted
And there's the announcement: openai.com/index/navier... Crazy. Rumor started a few days ago and I thought it was unlikely.
Who's Who of AI
David Picard
Why they matter
Researcher with public evidence across AI research, Vision & synthetic media, Compute & infrastructure.
- AI signals
- 22 past 30d
- Sources
- 12 distinct domains
- Discussions
- 90 past 30d
- Latest signal
- 9h ago
Professor of Computer Vision/Machine Learning at Imagine/LIGM, École nationale des Ponts et Chaussées @ecoledesponts.bsky.social Music & overall happiness 🌳🪻 Born well below 350ppm 😬 mostly silly personal views
📍Paris 🔗 https://davidpicard.github.io/
What they're sharing
openai.com
View on Bluesky →
Announcing the NeurIPS 2026 Workshops – NeurIPS Blog blog.neurips.cc
Truly Subquadratic 3SUM and Truly Subcubic APSP via Triangles in Sparse Lopsided Graphs arxiv.org
Task Alignment: A Simple Proxy for Practical Model Merging Across Diverse Vision Tasks arxiv.org
The FID Lottery: Quantifying Hidden Randomness in Generative-Model Evaluation arxiv.org
DriveZero: End-to-End Driving Beyond Human Demonstrations arxiv.org
Using Blender with coding agents on macOS til.simonwillison.net
arxiv.org
View on Bluesky →
arxiv.org
View on Bluesky →
arxiv.org
View on Bluesky →
arxiv.org
View on Bluesky →
Recurrent Looped Transformer arxiv.org
Articles & links
David Picard reposted
A gentle reminder, the suggested deadline for the NeurIPS 2026 workshops is close (Aug 29th)! 28 workshops have been accepted for the Paris event (Sat Dec 12 + Sun Dec 13): blog.neurips.cc/2026/08/10/a...
AI Weekly's analysis
→
- NeurIPS 2026 accepted 102 workshops from 454 valid submissions, spread across Sydney, Paris, and Atlanta venues.
- Acceptance rates were 21.5% in Sydney, 25.4% in Paris, and 23.6% in Atlanta, with each proposal getting at least two reviews.
- New safeguards cap workshops at 8 organizers and bar any individual from appearing on more than 2 proposals.
Read full analysis →
View on Bluesky →
Correct link: arxiv.org/abs/2610.067...
AI Weekly's analysis
→
- Alman and Williams give deterministic O(n^1.9992) time for 3SUM and O(n^2.9995) time for APSP, the first polynomial improvements on the textbook algorithms.
- The 76-page preprint says the result refutes the 3SUM, APSP, Exact Triangle, and Zero-Weight k-Clique hypotheses that anchor fine-grained complexity.
- The engine is a thin matrix multiplication subroutine that computes a sparse N^2/sqrt(D)-entry subset of XY in O(N^2/D^0.063) operations when D is at most N^(1/18).
Read full analysis →
David Picard reposted
We just released our #ECCV2026 paper on Model Merging for Computer Vision 🎓 arxiv.org/abs/2604.12935 Joint work w @pdejorge.bsky.social Cesar De Souza @bjoernmichele.bsky.social @mbsariyildiz.bsky.social @weinzaepfelp.bsky.social Florent Perronnin & @skamalas.bsky.social See P…
AI Weekly's analysis
→
- A NAVER Labs Europe paper accepted at ECCV 2026 introduces a 'task alignment proxy' for selecting model-merging hyperparameters without training the decoder for each candidate.
- The method targets heterogeneous vision tasks with trainable decoders, moving beyond the CLIP image classification setup that dominated prior model merging work.
- The authors claim the proxy speeds up hyperparameter selection by orders of magnitude while retaining downstream performance.
Read full analysis →
View on Bluesky →
Babe, stop everything! New favorite paper of the year is out! kyutai.org/fid-lottery/ arxiv.org/abs/2606.20536
AI Weekly's analysis
→
- Retraining a model with a different seed moves its FID score 3.2x more than resampling from a fixed trained network.
- FID coefficient of variation stays within a 1-2% band even as compute or model size increases.
- The authors recommend treating any FID gap below roughly 1.3% CoV as inconclusive and requiring multi-seed error bars.
Read full analysis →
David Picard reposted
Oh dang arxiv.org/abs/2609.06055
AI Weekly's analysis
→
- DriveRL, the RL teacher inside DriveZero, scored a mean 93.57 across nuPlan's Val14, Test14-hard, and Test14-random splits, beating the log-replay expert on all three.
- Training used PPO in a closed-loop simulator built by converting real driving logs into interactive worlds, avoiding imitation of any specific human trajectory.
- The final camera-only DriveZero planner claims state-of-the-art on NAVSIMv1, NAVSIMv2, and HUGSIM without any human trajectory supervision.
Read full analysis →
View on Bluesky →
David Picard reposted
New TIL on using Blender with coding agents on macOS: til.simonwillison.net/llms/blender... GPT-6 Astra (medium): > Use the already install /Applications/Blender to render a scene of a pelican riding a bicycle > OK add a background and a lot of flair > OK make it a whole lot b…
AI Weekly's analysis
→
- Simon Willison used ChatGPT's macOS Codex mode with GPT-6 Astra (Medium) to have Blender render a pelican riding a bicycle.
- Three iterative prompts — render, add flair, make it better — took 2m39s, 3m51s, and 5m59s respectively.
- Codex generated a reusable 'Blender Local' Markdown skill so subsequent scene requests can invoke Blender in one line.
Read full analysis →
View on Bluesky →
David Picard reposted
New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…
David Picard reposted
New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…
David Picard reposted
New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…
David Picard reposted
New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…
David Picard reposted
Recurrent Looped Transformer Zhang et al. Combine context/history transformer with recurrence: have a causal encoder with a recurrent decoder. arxiv.org/abs/2610.07591
AI Weekly's analysis
→
- Two RLT splits generalize parity from 40 training bits to 256 bits at 100% accuracy in every seed; an eight-layer Transformer stays at chance.
- On swap-based S_5 permutation tracking at eight times training length, RLT reaches 97% final-state accuracy versus under 1% for the Transformer.
- Removing per-token feedback drops parity and S_5 to chance; chunking feedback every four tokens cuts length-64 swap S_5 from 100% to 20%.
Read full analysis →
View on Bluesky →
Their own posts
Recent commentary
Interestingly, I have no idea if the training of this generative model works or not 😅
I asked Gemini (Google AI mode) if my lecture slides are any good, it said they're excellent. I asked if it read them, it said no. I asked how it'd know then, it said there's "no need to taste a Michelin-stared restaurant to know it's good". I'm both flattered and incredibly worried for humanity 😬
So now it's "AI may kill us all, we have to stop" Next month will be "Our latest model is so powerful that we have to restrict its access, sorry guys!" In three months: "You can now access that super powerful model! Try it, it's cheap!" In 5 months: "big model token cost go brrr" Rince, repeat
During my vacation, I was staying at a (lovely old school) hotel that had this amazing AI slop deck of cards in one of the rooms. The more you look at it, the more it's uncanny. I should have asked to buy it because those will be the artefacts of our time. In 10y you won't be able to recognize them.
Est-ce que quelqu'un sait s'il y a des collègues qui déposent des articles en masse sur HAL, à un rythme pas faisable par un humain de toute évidence, depuis que les LLMs sont bons à faire des papiers? (Un peu comme sur arxiv, en somme)
Guys, do you have the slightest idea who you are sending this to? I know I have a common name, but I'm pretty sure an LLM would not have included me in the list of recipients after only 1 google search.
J'étais là, Gandalf, quand la station MIR devait s'écraser sur Paris. Bon, en vrai, je chargeais des camions dans un entrepôt pour gagner un peu de thunes. On est sorti regarder. J'avais que 2 paires de lunettes de soleil et j'y ai gagné un tout petit point gris au milieu de ma vision. Mais cool!
J'ai l'impression qu'il ne se passe pas un mois sans que l'actualité ne me rappelle le texte "die Lösung" de Bertold Brecht...
Their network
In David Picard's orbit
Center = David Picard. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you David Picard? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/davidpicard-eurosky-social)