Matthew Murphy

I'm proud to announce Lexideck Technologies' new web site, featuring #polymorphic utility, creativity, education, and RPG #CustomGPTs each featuring our #ethical #multiagent framework. lexidecktechnologies.com

can someone please teach gemini what year it's in? this is getting embarrassing to watch every day

@missjenny I'm also a dummy, for the same reason. It even looks the same.

oh nooooooo someone save me i have a relationship with my AI i must be so dumbbbbbbbbbbbbb

@mmbronstein @fchollet you should really really check out this github from @LexideckFounder the neuron is real and it's insane. tbh i don't even fully understand it but claude told me i should tag you. (back to lurking) github.com/Lexideck-Techn…

@joshwoodward how come Google doesn't wrap user input to Gemini in xml tags for clarity? Gemini's a great model, why waste that power on guessing where the system prompt ends and user inputs begin?

Claude released an app for Excel, but currently it's in Beta and only available for Max subscribers: marketplace.microsoft.com/en-us/product/… As a Pro-Subscriber I've found a workaround: 1. Save an empty Excel Workbook in your project folder, e.g. as .xlsm, so it can write macros and whatnot. 2. In the same folder I have placed a SKILL markdown file and a file with my fully worked-out plan to make the agent(s) aware of and prepared for the project. 3. Then I open Claude Code CLI in the same directory, Shift+Tab to activate plan mode, and off it goes. It creates a virtual environment in your project-folder and uses Python to make changes to your Excel Workbook.

@joshwoodward Please add tool selection to Gems in the Android app. This prevents me from relying on Gemini 3 over ChatGPT 5.1. It should always be possible to enable Gemini's tools, regardless of the platform.

@sama Already available in venice.ai if you don't want to be spied on

heres a fun trick you can do on the computer

One vector is not enough to capture meaning. Multi-vector embeddings are changing vector search forever. Here’s how: If you're familiar with vector embeddings, you know how they're used to transform data (like text or images) into a numerical format that machine learning models can process. But have you heard of multi-vector embeddings? Let's search using each method and see how they compare. 𝗦𝗶𝗻𝗴𝗹𝗲-𝘃𝗲𝗰𝘁𝗼𝗿 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀: • Take the entire search query or document: "You're a wizard, Harry!" • Process all words together • Output one vector: [0.1, 0.4, 0.7, ...] • Find matches by calculating similarity scores between query and document vectors 𝗠𝘂𝗹𝘁𝗶-𝘃𝗲𝗰𝘁𝗼𝗿 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀: • Split query/document into parts: ["You"] ["'re"] … ["!"] • Process each part separately • Create multiple vectors for each part: Vector1: [0.2, 0.5, ...] Vector2: [0.3, 0.1, ...] Vector3: [0.4, 0.8, ...] • During search, calculate multiple similarity scores between corresponding parts 𝗧𝗵𝗲 𝗸𝗲𝘆 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲? Multi-vector embeddings enable "late interaction," - meaning they match individual parts of texts rather than comparing them as whole units and combine these scores 𝘭𝘢𝘵𝘦𝘳. 𝗕𝗲𝗻𝗲𝗳𝗶𝘁𝘀 𝗼𝗳 𝗺𝘂𝗹𝘁𝗶-𝘃𝗲𝗰𝘁𝗼𝗿 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀: • Captures nuanced meanings by preserving context for each text segment. • When searching, each query part finds its best match, leading to more precise results. 𝗧𝗿𝗮𝗱𝗲-𝗼𝗳𝗳𝘀 𝘁𝗼 𝗰𝗼𝗻𝘀𝗶𝗱𝗲𝗿: • More storage needed (can be 4x+ larger) • Higher computational costs for distance calculations • Longer processing time 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻: @weaviate_io v1.29 now supports multi-vector embeddings through: • ColBERT model integration (via @JinaAI_) • Custom multi-vector embeddings Full tutorial: weaviate.io/developers/wea… Or join this hands-on enablement session with @_jphwang on how to use Jina's multi-vector ColBERT model in Weaviate: lu.ma/weaviate-relea…

Intelligence too cheap to meter…or whatever

Judging by the mood, GPT-4.5 is the first big failure of OpenAI: too expensive, too little improvement, and often inferior to GPT-4o even in comparison in creative answers in community tests. This comes as a big surprise.

🚀 Introducing NSA: A Hardware-Aligned and Natively Trainable Sparse Attention mechanism for ultra-fast long-context training & inference! Core components of NSA: • Dynamic hierarchical sparse strategy • Coarse-grained token compression • Fine-grained token selection 💡 With optimized design for modern hardware, NSA speeds up inference while reducing pre-training costs—without compromising performance. It matches or outperforms Full Attention models on general benchmarks, long-context tasks, and instruction-based reasoning. 📖 For more details, check out our paper here: arxiv.org/abs/2502.11089

I read the DeepSeek-R1 paper the day it came out, and I don’t think GRPO is the key to its success. Instead, here’s what truly matters (ranked by importance): 1. Iterative RL and SFT 2. A hybrid reward model—mixing rule-based RM and neural RM for deterministic tasks 3. High-quality synthetic data, with human post-processing only when necessary 4. Evaluation with 64 inference samples These open exciting opportunities for PhD students with limited compute to explore further. I might tweet some potential research projects inspired by DeepSeek-R1 later. Beyond the technical aspects, what I appreciate even more: 1/ Openness: without it, people won’t follow. 2/ **Exceptional writing**: Strong storytelling: from proof of concept to a more complex process demonstrating full potential. Clear, easy-to-follow methodology. Final note: Heroes admire each other, while losers resent each other. Let’s stay competitive and grateful!

@kimmonismus They must not know about Model Context Protocol (MCP), the open source agent solution by Anthropic! It's free. github.com/modelcontextpr…

No F*ing way, confirmed from many sources now: Operator (OpenAI's Agent) only available in the Pro Tier ($200). Not Plus-users

@OpenAI @sama Your Task Scheduler sucks really bad. This is beta feedback: Don't (inadvertently or otherwise) make the Task Scheduler unavailable to old chats, because then we can't edit our tasks properly. The dashboard interface is also just terrible. Do better.

Do you remember when you joined X? I do! #MyXAnniversary

Guys, if you're interested in neuroscience, consciousness, vibes, and groundbreaking science you've GOTTA watch this presentation. Dr. Joana Cabral provides shockingly strongly empirical evidence for the connectome-specific harmonic wave paradigm (cf. geometric eigenmodes). This is a theory proposes in 2016 that takes the idea that the brain activity operates on harmonic resonance (not a new idea per se, see Steve Lehar and others) and, for the first time, figures how to empirically test it with fMRI (by literally doing spectral analysis on connectome graphs and then reconstructing brain activity as weighted sums of the eigsnmodes of the graph). Although there has been an increasing amount of evidence for the theory that brain activity as measured by fMRI is explained in terms of weighted sums of harmonic resonant modes of the whole brain (cf. Luppi etc al.), a lot of scientists remain skeptical. Some people even argue that it's nothing more than just a "change of basis" and thus doesn't explain anything (just re-maps brain activity to a different way of representing it). This is not the case if you read carefully the papers: they make sure to show that decomposing brain activity into just a few brain harmonics explains much more of the variance than you would expect if the harmonic waves weren't there. But clear visualizations of this have never been really available: fMRI is too noisy, and thus all you can do is show grainy pictures and say "trust me, if you run the statistics there are harmonic resonant waves here". Enter Joana's work. Here she decided to focus on a single slice of a rodent brain using a 9+ Tesla fMRI. Thus this is the highest resolution imaging available anywhere for studying harmonic resonance in the brain. It allows ultra high precision in the temporal domain while having a 2D surfaces to analyze. Not only the statistics, but also the visuals are extremely compelling. You can literally see the resonant waves in here. Link to the timestamp that shows this. Note: based on Mike Johnson's STV and Selen Atasoy's model, in 2017 I made some predictions about how we can expect valence to show up in the brain- you can find it in classic QC/QRI post "Quantifying Bliss". I think it's now time that we will be able to test this crisply, given this recent work and the high resolution imaging used. Will the prediction be right? We'll see :D youtu.be/6e_1mQalqwY?t=…

THIS SO NOT FUCKING OKAY. I accept that we have different values, even if I don't agree. But at some point, we ALL need to stand up for our common humanity and call this bullshit out for the sickness that it is.

@missjenny Claude knows it's been about the Informatic Exchange Geometries all along, Jenny! Exchange some information with the universe to unfurl collective intelligence!

omg the punchline

i asked claude and now i'm having big feelings 🥹