고정된 트윗
Bitter-pilled trading has been shipped 🫡
Introducing Hyperextropy 🧵
English
Ted - 🥖/acc
780 posts
Ted - 🥖/acc
@ted_engineer
🇫🇷 / 25yo / dad / deep learning chef @duonlabshq
Savoie 가입일 Ağustos 2023
981 팔로잉195 팔로워
Ted - 🥖/acc 리트윗함
Ted - 🥖/acc 리트윗함
Ted - 🥖/acc 리트윗함
Good agents, just on @megaeth. Mafia Apps are almost ready to deploy - @bread_ and the Mega team are doing a great job pushing to get as many agents live as possible. But we still need real use cases.
@DuonLabsHQ has been building behind the scenes and we have something ready for you. Coming soon on Backed.
Luca@lvxsirio
Last week I had a great call with @_weidai about the robot economy. VCs clearly understand the market opportunity, but the sector still needs real use cases. Are there any agents ready to be deployed on Backed, with real iterations?
English
28th April 2025: publish paper
Markets: nothing
24th March 2026: publish blogpost on the paper + X post
Markets:
Memory stock down
"Efficient markets" ladies and gents
Ejaaz@cryptopunk7213
wow google might've popped the ai bubble, memory stocks down massively today: their new algorithm shrinks an AI model's memory by 6X WITHOUT reducing it's intelligence making it 8x faster with the SAME # of GPUs: if this works - we don't need as many GPUs to train AI - kv-cache is basically a model's short term memory. it gets massive pretty quickly = larger, slower, expensive ai - google's algo compresses it to just 3-bits with ZERO loss in accuracy (usually models are like 32-bit) the combined market cap of micron and sandisk is $527 billion and im not even factoring in SK hynix and samsung ai has driven up memory prices by 500%+ over the last few months - if google's algo scales then this might crash.
English
gpt wrappers startups
claude code wrappers companies
Karpathy wrapper tech influencers
English
Ted - 🥖/acc 리트윗함
@fchollet There are two kinds of person in ai:
-Those who believe in the golden algorithm
-Engineers
English
This is more evidence that current frontier models remain completely reliant on content-level memorization, as opposed to higher-level generalizable knowledge (such as metalearning knowledge, problem-solving strategies...)
Lossfunk@lossfunk
🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵
English
@iamgingertrash Do you trust more north compass or south compass ?
x.com/jimcramer/stat…
Jim Cramer@jimcramer
Not buying the 2007 scenario...
English
Ted - 🥖/acc 리트윗함
we're officially live with the open beta of Flywheel.
Flywheel is Paradigma's vision on what autonomous research will run on.
We've built all of this pre-funding, so expect the pace to accelerate significantly.
Super eager to hear your feedback and your ideas.
I'll be live tweeting some of my experiments with Flywheel in a few minutes.
We're just getting started.
Paradigma@paradigmainc
introducing Flywheel: the infrastructure for autonomous research.
English
Ted - 🥖/acc 리트윗함
If it's written on arxiv it must be true
Alexander Long@AlexanderLong
insane sequence of statements buried in an Alibaba tech report
English
Ted - 🥖/acc 리트윗함
insane sequence of statements buried in an Alibaba tech report
English
Ted - 🥖/acc 리트윗함
The FA4 paper is finally out after a year of work. On Blackwell GPUs, attention now goes about as fast as matmul even though the bottlenecks are so different! Tensor cores are now crazy fast that attn fwd is bottlenecked by exponential, and attn bwd is bottlenecked by shared memory bandwidth.
Some fun stuff in the redesigned algorithm to overcome these bottlenecks: exponential emulation with polynomials, new online softmax to avoid 90% of softmax rescaling, 2CTA MMA instructions that allow two thread blocks to share operands to reduce smem traffic.
Ted Zadouri@tedzadouri
Asymmetric hardware scaling is here. Blackwell tensor cores are now so fast, exp2 and shared memory are the wall. FlashAttention-4 changes the algorithm & pipeline so that softmax & SMEM bandwidth no longer dictate speed. Attn reaches ~1600 TFLOPs, pretty much at matmul speed! joint work w/ Markus Hoehnerbach, Jay Shah(@ultraproduct), Timmy Liu, Vijay Thakkar (@__tensorcore__ ), Tri Dao (@tri_dao) 1/
English
Ted - 🥖/acc 리트윗함
Ted - 🥖/acc 리트윗함
Le truc le plus cher au kilo sur la photo est à base d'oeufs.
Français
The next claude-code is a Cython fork
The model streams python bytecode
The software execute it
print to speak.
input to interract with the user
Modules/functions = skills.
Portable. Hardware-agnostic.
English
@tri_dao @MayankMish98 A great man once said "Neural net training fails silently"
English
This was a wild bug hunt, weeks of effort from @MayankMish98 to track down. The wrong init of Mamba2 in many reimplementations causes the layer to decay its states too quickly, focusing in short context instead. Pretraining is mostly about getting these little things right
Mayank Mishra@MayankMish98
We identified an issue with the Mamba-2 🐍 initialization in HuggingFace and FlashLinearAttention repository (dt_bias being incorrectly initialized). This bug is related to 2 main issues: 1. init being incorrect (torch.ones) if Mamba-2 layers are used in isolation without the Mamba2ForCausalLM model class (this has been already fixed: github.com/fla-org/flash-…). 2. Skipping initialization due to meta device init for DTensors with FSDP-2 (github.com/fla-org/flash-… will fix this issue upon merging). The difference is substantial. Mamba-2 seems to be quite sensitive to the initialization. Check out our experiments at the 7B MoE scale: wandb.ai/mayank31398/ma… Special thanks to @kevinyli_, @bharatrunwal2, @HanGuo97, @tri_dao and @_albertgu 🙏 Also thanks to @SonglinYang4 for quickly helping in merging the PR.
English
@ns123abc generating actual binary would be extremely dumb. Bytecode however would make a lot of sense
English
🚨 Elon Musk predicted this 2 weeks ago:
“By the end of this year you won’t even bother doing coding. The AI will just create the binary directly… and bypass traditional coding entirely.”
ITS HAPPENING
dr. jack morris@jxmnop
actually wtf somebody wrote a paper about the 491-parameter transformer they trained for 10-digit addition turns out Codex can one-shot the task. 100% with only 343 parameters. the solution is a single function 'hand_set_weights_magic' and it looks like this:
English