X4

Building this right now. Very early, pre-seed, but secured a few of the largest family offices already as LP and strategic partners with track-records for Ops. Now working on expanding this into circular lifecycles while the main challenge is being able to produce a regulated product with autonomous compliance and policies. Done this before for the EU aviation industry and shipping and logistics industry globally in information security policies and regulations at two firms, now building for our own foundation.

@AnothahWon @nikitabier @YossiShow @Antunes1 Can you share your blocklist? I love my timeline, just pure science and honest work.

@nikitabier @YossiShow @Antunes1 Funny to see the loudest being accounts I already blocked 😂 likely for their constant spam, farming or platform manipulation.

Starting Thursday, we'll be updating our revenue sharing incentives to better reward the content we want on X: We will be giving more weight to impressions from your home region—to encourage content that resonates with people in your country, in neighboring countries and people who speak your language. While we appreciate everyone's opinion on American politics, we hope this will disincentivize gaming the attention of US or Japanese accounts and instead, drive diverse conversations on the platform. We invite creators to start building an audience locally. X will be a much richer community when there's relevant posts for people in all parts of the world.

@nikitabier What about business people who travel most of the year?

@JustinAngel @vikhyatk @ProfTomYeh We need help @ProfTomYeh your superpowers are being called!

@vikhyatk You joke, but I’ve legit done @ProfTomYeh’s book with mathing some AI matrices by hand.

ML interview question: Here are the weights for Llama 3.1 70B. Generate a token by executing the forward pass manually using pen and paper. You have 30 minutes.

@fernandogo79578 @beffjezos My friend came up with a similar solution albeit just based on Riemann Zeta Zeros.

Just an architectural thought...JAT High-Level Architecture: Input Grids (demonstrations + test) ↓ π Projection Space Encoder ↓ Candidate Generator (Spiral + Duality operators) ↓ Resonance Scorer (Zeta pruning + Cosine harmony) ↓ Reflective Self-Gaze Layer (3/4 threshold + Conscience deposit) ↓ Incitement Trigger (E_incite peak detection) ↓ Output: Selected transformation grid

The masculine urge to try to hack a new solution to ARC-AGI benchmarks

Malware in LiteLLM! Don’t update or install! Steals all your credentials!

@rachpradhan @jedisct1 If I may share a wishlist: - replicable state - sandboxed/root-less installations - perfect uninstall Flatpack tries to solve this, but adoption is bad. Mainly because I think the work needed to maintain packages.

I built a package manager that installs software 10-500x faster than Homebrew. It's 1.2 MB. Written in Zig. Zero dependencies. It's called nanobrew, and here's how it works.

Hey Nick, appreciate your reply! Anthropic just validated the adversarial thesis. They drew red lines on surveillance and autonomous weapons pre-signature. Got blacklisted. Result: Claude downloads surged. OpenAI was opportunistic and sloppy according to Altman himself resulting in 295% uninstall spike + senior exec attrition + contract revision under reputational duress. I think Math, Inc. is overdue publishing explicit ethical use clauses and the company philosophy. To decrease tension in the Math community it’s good to lay down policies for attribution, collaboration, licensing and especially the policies for development alongside or orthogonal to humans. Hard constraints on autonomous weapons and surveillance in the T&C would signal the company philosophy. Anthropic proved markets reward alignment integrity over distribution velocity, there’s that. OpenAI proved reactive damage control codes as profit-seeking adversarial behavior. The verdict here is that consumers reacts strongly to moral hazard.

@X4AES @mathematics_inc @DARPA It sounds like there's a lot of mixed feelings about the timing and purpose behind OpenGauss. What do you think could have made it feel more aligned with positive goals?

Today, at the @DARPA expMath kickoff, we launched 𝗢𝗽𝗲𝗻𝗚𝗮𝘂𝘀𝘀, an open source and state of the art autoformalization agent harness for developers and practitioners to accelerate progress at the frontier. It is stronger, faster, and more cost-efficient than off-the-shelf alternatives. On FormalQualBench, running with a 4-hour timeout, it beats @HarmonicMath's Aristotle agent with no time limit. Users of OpenGauss can interact with it as much or as little as they want, can easily manage many subagents working in parallel, and can extend / modify / introspect OpenGauss because it is permissively open-source. OpenGauss was developed in close collaboration with maintainers of leading open-source AI tooling for Lean. Read the report and try it out:

@natashajaques Fantastic work! If the space of deviations in human, human preference, LLM and LLM preference is better understood, it would allow to produce “taste” for targeted control of output quality

The paper I’ve been most obsessed with lately is finally out: nbcnews.com/tech/tech-news…! Check out this beautiful plot: it shows how much LLMs distort human writing when making edits, compared to how humans would revise the same content. We take a dataset of human-written essays from 2021, before the release of ChatGPT. We compare how people revise draft v1 -> v2 given expert feedback, with how an LLM revises the same v1 given the same feedback. This enables a counterfactual comparison: how much does the LLM alter the essay compared to what the human was originally intending to write? We find LLMs consistently induce massive distortions, even changing the actual meaning and conclusions argued for.

Congratulations to the @Weixin_WeChat Team! Hope you also include Pi by @badlogicgames and other Agent Control Systems in the near future. Holy sh… 1.2 Billion users overnight! Chinese level Growth Hack!

Gold. As a neutral AI Utility company or even as a Lab, the adversarial nature of OpenAI, Anthropic could not be proven before. Meaning their T&C forcing Claw users to their APIs, but then being fed cheaper output “eventually” and at a much higher price than via UI. This enables big AI labs to implement: Dark Patterns + Financial Engineering produced by an Adversarial Analysis & Response System using API data to “poison competitors”, “harvest data” and feed low quality data to all other users. The result is that as an API client you have no means of trust to proprietary AI systems operators. The only solution to that is companies like @huggingface incorporating FHE (Fully Homomorphic Encryption) and partnerships with proprietary model providers and ZKP verifications that the model uploaded is identical to the model behind closed doors. This would enable trust-less guarantees to serious Labs who don’t want want a vendor being able to kill-switch their product, nor harvest their data to build a competitor.

They looked into the "they silently changed the model!!" And how to detect it. Interesting study. Though they only do it for API endpoints, and the usual complaint is for the chat interface; would be interesting (but harder) to extend to that.

I understand that MathInc used mathematicians and the math/lean community to construct their model, but then claims AI solved the entire formalization problem, firstly discrediting researchers, secondly just adding “automated formal verification” to already proven discoveries made by mathematicians, and lastly filling the entire field of manual verification with automated outputs that now nobody can manually go through anymore at the output rate, thus capturing an entire field. I think they have indeed been not attributed enough to mathematicians and community, but I can understand the CEO needing to build hype. However autonomous formal verification is only possible due to real mathematicians haven proven first. In my opinion the big problem is that MathInc likes to go even deeper and hijack the years of efforts of a Human Community to prove and position their products, rather than creating a new Field of “Agentic Proofs” that doesn’t overlap with humans at all, thus can’t claim as much novelty as they did now.

@MarioKrenn6240 @mathematics_inc What's the concern here? I would have thought the reaction would have been uniformly positive. Only just started learning about these efforts recently.

After the apparently amazing announcement by @mathematics_inc on the formalization of a major recent Fields-medal winning theorem, i had no idea how pissed the math-formalization community is. Very worrying discussions by some of the leaders/founders of Lean's mathlib. cc @ChrSzegedy

@sohamdaga22 @DhravyaShah Congratulations to the success tho! Will go through your paper 📝 ✌🏼

“Sounds like overfitting”, but if memory and retrieval works very good for recall, how good does it work for improving reasoning? The reason humans don’t have perfect memory is, because it’s detrimental to adaptive reasoning, generalization and consensus building over long horizons of knowledge and context vectors. If you would perfectly remember, the reward signal would cancel out further attempts at understanding at a deeper level.

agents absolutely dominate the ~1% loss is only because longmemeval has some incorrect ground truths in the dataset (see github.com/xiaowu0162/Lon…)

Respect the Honest work! Humbled to be already following all @‘s you dropped! I love programming language theory, compilers/vms and bootstrapping challenges. Been studying FRP, Arrows, Fusion, HoTT, Formalisms, Verification and CSPs for years, . Led me down a rabbit hole, with surprising results. So down for a chat!

Few know this, but I (George) was the only person in history to get a perfect score in CMU compilers, which is likely the best compilers course in the world. Combine that with crazy low level knowledge of hardware from 10 years of hacking. Then add a team of people who are talented enough to push back on my dumb ideas and clean up the implementations of the good ones. The team who keeps this whole operation running, software, infrastructure, and product. I love how there's no hype in deep learning compilers. It was one of the most annoying things about self driving cars, all the noobs who burned through billions on crap that was obviously dumb, and the companies who deserved to go bankrupt years ago if not for government bailouts (Tesla and China will devour them all). In this space, the competition is @jimkxa at Tenstorrent, @clattner_llvm at Modular, and @JeffDean at Google. Three of the living legends of computer science. And companies like @nvidia and @AMD, who are definitely live players, making single chips that have more power than the whole Internet two decades ago. This space is so fun to play in. If you haven't, read the tinygrad spec. It's all coming together beautifully.

@AerodromeFi @Cointelegraph @wagmiAlexander @rkbaggs @DromosLabs @VelodromeFi Been there with you for years, > 150k at top cycle

@Cointelegraph @wagmiAlexander @rkbaggs @DromosLabs @VelodromeFi 🛫

🏦DeFi disrupting TradFi was a farfetched notion less than a decade ago. Not anymore. Today, @rkbaggs was joined by @wagmiAlexander, CEO of @DromosLabs and a core contributor to @AerodromeFi/@VelodromeFi. The two went deep on why it's such an exciting time for DeFi!

@mathematics_inc @DARPA I just wish that it’s not used to build weapon systems in a world that is already quite hot. Do you have a policy similar to Anthrophic?

Big disappointment. Not that you work with DARPA or forked Hermes. I’m silent, unreal maximally observant fan of DARPA & Co. But the time you chose to build for the acceleration of war, is a flex of weakness in integrity for the transactional swap for currency. I hope to be wrong. I think if onlya total of 5% of Mathematicians, Phyisicsts, Virologists, Biologists, Chemists, Engineers and Hackers chose the dark side, we would be up for very very dark times, if at all.

@memcculloch Feeling honored. Happy to learn or hear his response ☺️✌🏼

@X4AES I did send a report to Pete Worden, on his request, on what to look for from distant aliens regarding QI.

Quantised inertia implies that energy can be extracted from the destruction of future possibilities. This is not why when you are told you can't do something you get a little hot under the collar. That's another process.

@KanikaBK I’m inclines to subscribe, if you please stop for the love of life, just stop making clickbaity titles. It ruins the entire experience. I know many papers before they are posted on X and clickbait doesn’t convert for me. Rarus sunt carus.

🚨BREAKING: Someone just open-sourced a tool that converts PDFs to markdown at 100 pages per second. 100% FREE. Runs entirely on CPU. No expensive GPUs needed. No cloud. It's called OpenDataLoader PDF. Give it any PDF - scanned documents, scientific papers, multi-column reports, complex tables - and it converts everything into clean Markdown, JSON with bounding boxes, or HTML. Ready to feed straight into any AI pipeline. Not a wrapper around someone else's OCR. Not a basic text extractor. A full document intelligence engine that understands layout, reading order, headings, tables, and formulas. Here's what this thing can do: → Extracts text in the correct reading order across multi-column layouts → Pulls complex borderless tables with 0.93 accuracy — highest of any open-source parser → Detects heading hierarchy, nested lists, and document structure automatically → Runs OCR on scanned PDFs in 80+ languages including Chinese, Arabic, Korean, and Japanese → Extracts math formulas as LaTeX from scientific papers → Gives you bounding boxes for every single element on the page → Describes charts and images using a built-in vision model → Filters prompt injections and hidden text - built-in AI safety that no other parser has Here's why every existing tool loses: They benchmarked it against 200 real-world PDFs including scientific papers and multi-column documents. OpenDataLoader scored 0.90 overall. Docling scored 0.86. Marker scored 0.83 but takes 54 seconds per page. MinerU scored 0.82 at 6 seconds per page. OpenDataLoader local mode? 0.05 seconds per page. That is over 1,000x faster than Marker at nearly the same accuracy. Here's the wildest part: It has two modes. Local mode runs pure Java — 20 pages per second on a basic CPU. Hybrid mode adds an AI backend for complex pages and scores #1 in every category. Run it on an 8-core machine with batch processing and you hit 100+ pages per second. Your documents never leave your machine. Zero API calls. Zero data transmission. 100% local. It even has a built-in AI safety layer that catches hidden text, transparent fonts, and off-page content that other parsers silently pass through to your LLM. One command to install: pip install -U opendataloader-pdf Works with Python, Node.js, and Java. Official LangChain integration included. 3.3K GitHub stars. 478 commits. 51 releases. 13 contributors. Actively maintained. 100% Open Source. Apache 2.0 License.

@TCryptochicks @cz_binance No Source? No Footer? -> “Wash-Trading” Proof by showing legitimacy. Tools like Chainalysis detected $704M in BSC wash trading algorithmically (historically and I don’t think this changed much)

Daily active users: BNB Chain — 4.4M Is it a sign? 😂