jeseem

@Yuchenj_UW It’s a problem everywhere in Bay Area. Almost no café open in evening.

Guys, SF is a magical city. Today, a beautiful Wednesday, I had a 4:30pm meeting with a friend at a café… it was closed. We walked 10 blocks, every café was closed. Finally found Blue Bottle, it closed at 5:30. Can some YC company please build what SF people actually want?

@mitchellh why not use both ? I use one to code and the other to review, Opus 4.6 to code and Codex 5.3 to review and then switch at times. It really helps and the review model catches lots of issues.

I know this is pretty well established at this point, but Codex 5.3 is a much more effective model than Opus 4.6. I went back and forth on both for a bit, but haven’t touched Opus at all now for a full week. First model to get me off of Opus… ever. Good job Codex team.

@karpathy languages that llm are good at will thrive, not what developers are good at. LLMs will prefer more concise language, to keep attention window short. And languages that inherently protect against coding errors and where compilers catch a lot of issues.

I think it must be a very interesting time to be in programming languages and formal methods because LLMs change the whole constraints landscape of software completely. Hints of this can already be seen, e.g. in the rising momentum behind porting C to Rust or the growing interest in upgrading legacy code bases in COBOL or etc. In particular, LLMs are *especially* good at translation compared to de-novo generation because 1) the original code base acts as a kind of highly detailed prompt, and 2) as a reference to write concrete tests with respect to. That said, even Rust is nowhere near optimal for LLMs as a target language. What kind of language is optimal? What concessions (if any) are still carved out for humans? Incredibly interesting new questions and opportunities. It feels likely that we'll end up re-writing large fractions of all software ever written many times over.

how much time do you spend in claude or codex planning phase ? #Claude #codex

@deedydas Using Claude by itself, becomes a tool for faster coding or to some level automating coding. But there’s more work needed to put the code in context of the whole project.

I'm not saying *everything* is solved. But Opus 4.5 seems to be this giant step function from say ~60% to ~80% of tasks where even L5/6 engineers feel like they mostly check code, or reprompt. Their job is just prompting for 5-10min tasks in 3-4 git worktrees at a time.

A few software engineers at some of the best tech cos told me this week "My entire job these days is prompting Cursor or Claude Code with Opus 4.5 to do what I need and sanity checking it." We've crossed some intangible threshold of AI generalizing to "most" software.

@SemiAnalysis_ This is so spot-on. 50% MFU effectively means you are using only 50% of total max. capacity. In any other cloud service, no one will accept such performance. Improving MFU for large clusters is one of the most urgent work needed. It's not an easy problem but the rewards are clear

Everyone likes to compare chips using peak FLOPs, but the additional puzzle to the formula of Performance per TCO is Model Flops Utilization (MFU) MFU is the real-world metric that determines how much compute you actually extract per dollar. And MFU varies wildly across users, workloads, scaling sizes, and software maturity. Most people don’t realize how sensitive it is: small changes in kernels, sharding strategies, or batch-size decisions can swing utilization by 2–3×. MFU drops as clusters scale up because communication overhead grows faster than compute. A well-tuned single node might hit 40 to 50%, but at multi-thousand-GPU scale, diminishing returns take over. This is why even elite runs like Llama-3-405B only achieved ~43% BF16 MFU on 16,000 H100s (however this is for Dense Models, whereby MoE tends to add irregularity reducing MFU) , and why poorly tuned workloads routinely sit in the 10 to 20% range. Precision also matters: FP8 MFU is far lower than BF16 MFU, and early-life platforms also tend to have lower MFU figures before software matures. Big labs are better - OpenAI, Anthropic, Meta, Google - all have expert teams and custom kernels that push MFU higher than the average user.

@chelseabfinn thank you. was looking forward to watching this since I read about the stanford course

All of my Deep RL course lecture videos from Spring 2025 are now online! 🥳 Youtube playlist: youtube.com/watch?v=EvHRQh…

@arif_ahmad_py @anand_bhattad @jathushan @HaoLi81 @konpatp @GurushaJuneja This is a master class in how to present a poster session. It shows the experience in how the core idea is presented in its most basic form. Very interesting paper and amazing presentation skills

We need more senior researchers camping out at their posters like this. Managed to catch 10 minutes of Alyosha turning @anand_bhattad’s poster into a pop-up mini lecture. Extra spark after he spotted @jathushan. Other folks in the audience: @HaoLi81 @konpatp @GurushaJuneja.

@blelbach Was waiting to try this after seeing the last present at PyTorch conference. Is it supported in PyTorch also?

CUDA Tile has shipped! You can now `pip install cuda-tile`. I'm excited to see what y'all will build with it! Docs & resources: developer.nvidia.com/cuda/tile GitHub: github.com/NVIDIA/cutile-…

@GoogleResearch This has definitely captured one of human memory retention concept imho. Ad orgs probably have used these and similar ideas to capture human attention. Very interesting work

Today at #NeurIPS2025, we present Titans, a new architecture that combines the speed of RNNs with the performance of Transformers. It uses deep neural memory to learn in real-time, effectively scaling to contexts larger than 2 million tokens. More at: goo.gle/3Kd5ojF

@nouhadziri Congratulations

Beyond delighted about our #NeurIPS2025 Best Paper Award 🥳🥳😍😍🥇🥇🥇

@dylan522p the change to NL72 from 36 is interesting. Is there any work in offloading inter-rack traffic, aggregating at rack level, like my expanding Nvidia SHARP ?

Only NVIDIA, AWS, Google have successfully deployed rack scale architecture Trn3 is the 2nd after Nvidia with switched scale up topology, which is better for frontier mixture of experts models AWS Trainium3 course corrects on software Switch design choices are... interesting

Some interesting presentations and paper talk at NeurIPS. Some very relevant papers, directly relevant to me like the multi-turn agent reasoning. One of the work is training RL on dynamic state data and multi-turn agents. #NeurIPS2025

@chrmanning Hi Prof Manning, are you presenting any session at NeurIPS

I’m now in San Diego for NeurIPS too! 😆

@Mascobot @a16z this is the best dev server option. earlier this year, after struggling to find GPUs, I was very tempted to buy a 2-RTX6000 server. This is way better, especially for training smaller model or where the RL is slow as the env response is slow

Jensen came to our @a16z's Runtime event and he signed our very first personal GPU AI Workstation Founders Edition (4x RTX 6000 Pro Blackwell) "To a16z Builders of Tomorrow!" - Jensen

Attending the NeurIPS conference #NeurIPS2025

Recharging my and me at TeslaDiner on my way to NeurIPS

@yunwei37 thanks for the excellent blog post, something I was looking for.

A blog for GPU observability tools using bpftime eunomia.dev/blog/2025/10/1…

@GarrettDrinon The best swing imho was Metsera $MTSR. The entry was on Novo bid and the bidding war is still on.

continuing to be verrry selective with swing trades the volatility of market oscillations is telling me to wait sometimes it’s the action that carves out a top sometimes it’s just a pause could change on a dime but… my process tells me to shorten the timeframe, be patient, scalp out risk, and take the money when it’s there

@ezyang Uneven sharding and fsdp FSDP and cpu offloading and Device mesh; process group initialization having these would be great for the podcast

I've been brainstorming episodes for the next season of PyTorch Developer Podcast. DTensor StridedShard, FSDP-TP order Redistributing a DTensor Prefetching vs Bucketing History of FSDP in PyTorch Multiprocessing: DataParallel versus DistributedDataParallel Monarch Parallelism Zoo Mixture of Experts and Expert Parallelism The Peak Memory Triangle: Activations FSDP and CPU Offloading Overlap: How to get it (Prefetching, Pipelining, Async TP) Differentiable collectives and variance Local map: global versus local SPMD FSDP vs TP Symmetric memory Uneven sharding and FSDP LocalTensor Composable parallelism via DTensor Pipeline parallelism Functional collectives and wait Device mesh; process group initialization Distributed checkpointing Activation checkpointing Placement: Partial reductions Implicit versus explicit prefetching RNG in a distributed setting Distributed optimizers, Zero, Shampoo, Muon torchtitan torchrun / torchx Choosing your parallelism from first principles / roofline analysis Mixture of Experts: as large as possible, expert routing as fine as possible (more sparsity the better) by minimizing hidden dim GB200 MXFP8 (1x128, 128x128, transposes) Stats of a training job: loss curve, MFU, expert balance Bitwise determinism: when you can expect it Distributed inference RL from an infra perspective Basics of observabilty on jobs NCCL timeout