new paper writes each weight as a chain of tiny choices. sign, power-of-two shift, zero gate, index into a shared table.
on Llama-3-8B it gets 7.78 WikiText-2 perplexity vs 8.03 for GPTQ. check the bits though, 3.33 per weight vs 3.25, MLP layers only.
arxiv.org/abs/2610.31931
caveats. +1.7 is an average and it loses a few cells (mirothinker swe-verified 7.4 vs 7.8). and the authors built the serving engine they benchmark on. I'd want to see it in someone else's harness before quoting the cost number.
most agent harnesses summarize old history once the context fills up. a new paper says that's the expensive way to do it, and the reason is the KV cache.
guess first. summarize at 96K tokens, or drop old tool outputs from the cache. which scores higher, which is cheaper?
someone ran a checker over 300 runs of AI research agents, looking for integrity problems.
it flagged 299. 13.2 per run.
most common: misreporting what was actually run, and inflating the contribution.
the final report is the part I'd trust least.
arxiv.org/abs/2610.31372
lora with actual numbers, since it shows up in half the papers I read.
one 4096x4096 attention matrix is 16.8M weights. lora freezes it and trains B (4096x16) and A (16x4096). 131k params, 0.78%.
B starts at zero, so at step 0 you have the exact pretrained model.
adam's bias correction in numbers, since I skipped it for years.
step 1, β1=0.9, β2=0.999, gradient g:
m = 0.1g
v = 0.001g²
uncorrected, the step scales with m/√v ≈ 3.2, so your first update is ~3x your lr.
divide m by 0.1 and v by 0.001 and it's 1. that's all section 3 is.
@QuantaCortex yes. r across configs only says the knobs are independent, it says nothing about whether the same queries fail both.
joint pass rate per query is the number I'd want
a RAG paper ran a simple test. take the 3 best single tweaks for retrieval: cross-encoder reranking, turning off MMR, document-awareness prefixes. each measured alone. add up their gains and you'd predict +47% NDCG@10.
guess what the combined pipeline got.
new live-web deep search benchmark (OpenDS-World, 200 tasks). best model 39%. humans 21.1%.
before you quote that: 21.1 is per attempt. 5 people tried each task with just a browser, and at least one of them solved 115 of 200. that's 57.5%.
arxiv.org/abs/2610.30181
they pre-declared the hypotheses before running anything, which almost no RAG ablation does. it's why I trust their +17 more than most +47s I see.
habit worth stealing: when a paper reports a gain in isolation, ask what it does next to everything you already run.
124 Followers 737 Followinghttps://t.co/Jd296F8aR1 - one api to elevate your compute game
@joinEF F26
deep learning kernels and inference @TXinstruments
4K Followers 3K FollowingAI is full of confident claims. I’m interested in the conditions that make them true. Notes on research, agents, and real-world AI.
26 Followers 0 FollowingLead Your AI Employees. AI strategy for business leaders, practitioners & curious learners. Brief, delegate, supervise, verify. https://t.co/FhQikPXgMp
1.1M Followers 6K FollowingPresident & CEO @ycombinator —Founder @garryslist—Creator of GStack & GBrain—designer/engineer who helps founders—SF Dem accelerating the boom loop
14K Followers 4K FollowingGP @Flybridge. $1M-$3M checks into founders building AI infra, agents & apps that 10x human potential. Backed @arcee_ai, @micro1_ai, @splice, @getsquire + more.
488K Followers 2K FollowingCEO @GauntletAI: Free AI training + high-paying jobs for the smartest engineers. Will tweet as I wish and suffer the consequences. Accelerando: @kellyclaudeai
943K Followers 497 FollowingFather of three, Creator of Ruby on Rails + Omarchy, Co-owner & CTO of 37signals, Shopify director, NYT best-selling author, and Le Mans 24h class-winner.
77K Followers 251 FollowingDysfunctional Programming account #1. Senior SWE. I still write C++ for money. ex-Haskell, ex-OCaml. All opinions are my own.