Start with something like PopperPad.
PopperPad is scientific memory for humans and agents: an append-only, falsification-first knowledge ledger for hypotheses, recipes, evidence, counterexamples, artifacts, and typed relationships between claims.
github.com/TheDarkLightX/Pop…
6
706
A man who pushes open weight open ecosystems, is being called a tyrant and authoritarian, for running the most successful tech company in America?
1
16
Being a CEO is always about relationships.
很多人一直没看懂黄仁勋。
你以为他是硅谷最顶尖的科技狂人,其实老黄骨子里,就是一个极其老派的亚洲生意人。
他的生意经,中国老板一看就懂。
讲交情,重信用,认老朋友,关键时刻愿意亲自下场。技术要过硬,人情也得做到位。
英伟达最厉害的地方,除了CUDA和GPU,还有老黄花了几十年经营出来的亚洲供应链关系网。
你仔细看看,台积电在台湾,先进封装的重要产能在台湾,HBM内存的核心供应商在韩国。
这些企业背后,都是一群在半导体行业摸爬滚打几十年的老江湖。
老黄跟他们打交道,靠的不只是采购合同,还有几十年攒下来的交情、信用和相互成就。
尤其是台湾半导体圈,很多商业关系带着浓厚的传统商帮色彩。
合同当然要签,但合同解决的是法律问题,交情解决的是关键时刻谁愿意帮你。
假设全球产能都紧张,十家公司同时抢一条生产线。
一个职业经理人带着律师团队过去,拿出几百页合同,讨论价格、交期、违约责任。
老黄可能直接一个电话打给认识几十年的老朋友:
“兄弟,这次真得帮我一把。”
对方愿不愿意帮,当然还得看利益和产能,但几十年的信任,绝不是几张合同能够替代的。
这才是老黄最难复制的商业资产。
华尔街可以找一个懂财务的CEO,也可以找一个懂技术的CEO,甚至可以找一个比老黄更年轻、更会管理的CEO。
但上哪儿找一个能把整个亚洲半导体产业链经营了几十年的人?
你以为英伟达卖的是芯片。
其实老黄手里还有一张看不见的王牌:几十年积累下来的产业信用和私人关系。
硅谷那帮人研究怎么用AI改变世界,老黄却把中国商人最传统的一套生意经,做到了万亿美元级别。
科技可以迭代,合同可以重签。
唯独几十年的交情,没办法花钱买,也没办法一夜之间交接。
这才是黄仁勋真正难以替代的地方。
2
83
Replying to @thsottiaux
Less you want to concede Cyber to China, give mfs with proven track record Daybreak Red or Purple or whatever the fuck works for actually doing security work. Or don't. Idgf anymore.
You mfs trained on my peers work from the past 30+ years and only a handful of them have access. A few others have gotten 3-4k usd compensation and the rest can't even access a model without being told what they can or cannot do.
It's fucking ludicrous the nr of people I know from IC that don't have access.
D.#dwards retweeted
Replying to @HowToPrompt__
Grep earns its keep when one character in an error code changes the answer. Semantic similarity isn’t always the match you need.
D.#dwards retweeted
A free tool from 1974 just embarrassed the modern AI in this paper
It’s called Grep.
it's an old command line tool. It's so old that it predates the internet itself. It's basically a control + F.
Every AI startup right now is selling you on vector search.. embeddings, semantic search, chunking, rag systems..
But researchers at PWC just dropped a paper where they put modern vector search head-to-head against grep.
this was a massive study.
5 different models (gpt, claude, gemini) and 4 different agents digging through 116 questions across months of conversation history.
they asked specific questions like "what restaurant did someone mention 3 weeks ago?" or "what date did they travel?"
Grep won. in every single model.
when you think about the math.. it makes sense. vector search is designed to find things that are "conceptually close" to what you are asking.
which is great.. until close is just the wrong answer. grep doesn't do close. it finds the exact word or it finds nothing.
and for literal questions, exact matching destroyed millions of dollars of ai r&d.
but here is why this paper is actually kind of wild.. the whole test was basically rigged for grep to win.
every question in the test had a literal answer in the data.
but what if you remember someone feeling "stressed", but in the chat they actually used the words "overwhelmed" or "burnt out"?
grep will give you a blank screen. vector search will find it instantly. that’s literally why embeddings exist..
plus, we already have keyword search with bm25 for hybrid approaches in modern rag. going back to raw grep makes zero sense outside of this one hyper-specific use case where you know the exact word you are looking for.
i honestly don't know what pwc is doing with this paper..
let me know if i missed something.
All problem solving is search, and all of can be thought of as directed graph path finding. There is nothing shocking here. This is what I expected a model to do.
Great paper on how AI models learn to reason.
We all wonder how do LLMs actually learn to reason step-by-step instead of just memorizing patterns?
You can see how they use directed graph path-finding to finally give us a rigorous, mathematical peek under the hood of how these models think.
The math they laid out is just so elegant.
They proved that a 1-layer transformer is completely bottlenecked, needing a massive O(N^2) size just to figure out a basic child-selection step in a depth-first search.
But the second you add one more layer, everything changes.
It turns out a 2-layer model is completely sufficient to handle full DFS reasoning with a highly efficient O(N log(N)) size.
It gets even crazier as you scale up, because they showed that for anything three layers or deeper, the model’s look-ahead horizon actually grows exponentially with depth.
They proved these two-layer models can actually learn this DFS process through plain old gradient flow and perfectly generalize it to entirely unseen graphs.
It is incredibly rare to see theoretical training dynamics backed up this cleanly to explain why depth improves reasoning efficiency.
The fact that they dropped full Lean 4 proofs and open experiment code alongside it is just a massive flex for reproducibility.
Definitely worth reading this: arxiv.org/pdf/2611.13594
1
5
242
In the future, humanity is likely to be monitored 24/7, 100% of the time, by AI, as soon as technology makes it feasible. Do you still want data centers knowing why your neighbors want them? How do we know it would stop at murder and rape?
D.#dwards retweeted
While CRAM is not yet a general reasoning model, our work suggests that energy is an exciting avenue to explore as it offers huge speedups and data-efficiency over looped methods
40 years in, looks like physics had it right
D.#dwards retweeted
In fact, CRAM energies hand-coded with knowledge of the task can solve all 4 benchmarks, using no more than 3 parameters. This suggests that these common benchmarks are much too simple to really test reasoning
We need a better suite of basic reasoning tasks
D.#dwards retweeted
CRAM needs less data to train than looping models, reaching ~100% Sudoku-Extreme test accuracy with 30 puzzles, compared to 700+ puzzles for larger looped models
Learning constraints is a lot easier than learning reasoning trajectories
D.#dwards retweeted
CRAM does this with no loss of accuracy across four favourite benchmarks of recent reasoning models.
Moreover, energy tells you when to stop looping and when to keep going. This lets compute be scaled for harder problems.
D.#dwards retweeted
A perk of energies is that they can be trained without backprop-through-time (BPTT). Just pull down the energy of data samples, and push it up everywhere else. That is enough for a meaningful energy
Avoiding BPTT gives HUGE speed boosts for training
D.#dwards retweeted
We call this network CRAM: a Compact Reasoning Associative Memory. Prior works set the weights of these networks by hand using knowledge of the puzzle’s constraints. We ask if they can be learned.
They can. And the results are really good
D.#dwards retweeted
Looped models are expensive, unstable, and slow to train. For simple benchmarks like Sudoku and Maze, is there a better way?
Energy descent, using a trainable version of Hopfield & Tank’s energy from 1985: stable dynamics, early stopping, fast training, and more
D.#dwards retweeted
Big reasoning speedups are here. 1,000× faster training and 280× less compute than any looped model. All done using a fully trainable energy from the 80s
Keep reading to hear how energy is the right paradigm for reasoning
With @HughBlayney @jcrwhittington @prefrontal_ai
🧵/10
D.#dwards retweeted
Finally wrote a blog post about this! Normal LLM inference isn't deterministic across devices, but this can be fixed.
Floating-point addition isn't associative, (a + b) + c ≠ a + (b + c), and accumulation order changes with the hardware and kernel used, leading to different outputs on the same prompt.
Integer addition is associative, so doing every op in integers (including softmax and norms) gives bit-identical outputs across GPUs and CPUs.
Github repo and post listed in the comments
Turns out you can make LLM inference fully deterministic across devices, with no loss to quality or speed.
This weekend at the @spacexai hackathon I got Qwen3-0.6B to produce identical hashed logits from a 512-token generation across 2 GPUs and 3 CPUs: an A100, an H100, an Apple M5 Max, an AMD EPYC, and an Intel Xeon.
The main reason why inference isn't deterministic is because floating-point addition is not associative. (a + b) + c ≠ a + (b + c), since every add rounds. Accumulation order changes with the hardware used and kernel selected, so the same prompt can give you different outputs even at temperature 0.
Integers ARE associative. So why doesn't integer quantization already fix this? Because, while weights and activations get quantized, the non-linear ops (softmax, normalization, SiLU) dequantize back to float and requantize afterwards. Each of those steps hands you back to floating point rounding.
True integer-only inference does exist, but it's historically been motivated by edge hardware without FPUs, which doesn't make much sense for LLMs. One 2024 paper (I-LLM) did it on LLaMA from that angle and didn't get much attention. Nobody seems to have looked at it from the determinism side.
I wrote my own implementation, simplifying the approach from the paper, so that every operation between the input ids and the int32 logits is exact integer arithmetic. To test it I chain-hashed the logits at every step and ran that across the devices and configurations below. Every integer run gave the same hash: 64430dd985f8. Every fp16 run gave a different one, all diverging on the very first token.
WikiText2 perplexity came out to 20.72 vs 20.95 for fp16 (slightly better than the float baseline), and CUDA-graphed integer decode hits 106 tok/s at batch 1 on an A100, 3.6x the fp16 eager baseline.
Github repo is listed in the comments. Plan to do a writeup over this eventually!
Now we will see AI slow down. It's over.
JUST IN: *Nvidia ($NVDA) is expected to slow its aggressive AI investment strategy following Jensen Huang's resignation, putting billions in startup funding and credit guarantees at risk, The Information reports.
46





















