It’s fascinating how Democrats never call for George Soros to be taxed more. His net worth is $7.5 billion. https://t.co/r3UWxlB056

Got the model converted to CoreML and working on iOS; will open source soon! https://t.co/6xo8VetVGT

Got the model converted to CoreML and working on iOS; will open source soon! https://t.co/6xo8VetVGT

GLM-5.2 is now @Zai_org's most-liked model on Hugging Face of all time. https://t.co/qFgxVrzer8 https://t.co/SIS5y0BEN1

GLM-5.2 is now @Zai_org's most-liked model on Hugging Face of all time. https://t.co/qFgxVrzer8 https://t.co/SIS5y0BEN1

X just launched hosted MCP servers so AI tools can connect directly to the platform. Connect Grok Build, Cursor, Claude, VS Code, or any MCP client to two official servers: • X MCP (httpx://api.x.com/mcp) search posts, manage bookmarks, fetch trends/news, and draft/publish Articles with your account permissions. • Docs MCP (httpx://docs.x.com/mcp) instantly search and read X API docs inside your workflow. Quick Grok setup: X Developer Portal: httpx://developer.x.com/ Before you begin. Create an X app in the X Developer Portal with OAuth 2.0 enabled. Register the redirect URI https://localhost:8080/callback on the app. Copy your CLIENT_ID and CLIENT_SECRET. Replace YOUR_X_APP_CLIENT_ID and YOUR_X_APP_CLIENT_SECRET in the attached image with these. Run the commands from: [Image attached at bottom] Then verify: grok mcp doctor xapi grok mcp list One-time browser login on first run. Tokens cached and auto-refreshed locally after. Big unlock for agentic systems that need live X data.

This is a classic meme. But it's not the entire picture: 1) Most of the grift and graft happens at the NGO level with overpaid liberal midwits outsourcing to outsourcers, just so they can go to Davos and present themselves at cocktail parties. 2) None of these people have moved the needle in decades. 3) It's patently dishonest to blame Elon for the deaths of people in Liberia or anywhere else as a result of demanding accountability for how we spend the tax dollars of hardworking Americans. 4) If the solution is so simple, MacKenzie Scott, Laurene Powell Jobs (friend of Ghislaine Maxwell), or Nancy Walton (owner of a $300 million mega yacht) could snap their fingers and solve the problem. This entire anti-Elon, media driven narrative is so obtuse and contrived. Anyone who buys into this is intellectually captured. Foreign aid creates dependency and props up failure. Real progress comes from property rights, rule of law, trade, entrepreneurship, and stable governance, not blank checks. Stop repeating the same failed experiment and expecting different results.

One vote stood between America and independence. The man who had to cast it was 80 miles away, battling cancer, as a violent thunderstorm raged. His name was Caesar Rodney. A lesser-known founder of the American Revolution, Rodney rode overnight on horseback from Delaware to Philadelphia, arriving just in time to cast the deciding vote for independence and sign the Declaration of Independence. 🇺🇲

85% of reported scam complaints involving social media were linked to Facebook, while 𝕏 had the lowest at just 0.1%. • Facebook was cited in 85% of identified social media scam complaints collected. • 𝕏: 0.1%, the lowest among major platforms. • Internal Meta documents projected $16 billion in annual revenue from scam and banned-goods ads, about 10% of the company’s revenue. • The watchdog SafelyHQ has received 50,000+ verified scam reports, with experts saying the real victim count is likely in the tens of millions. • Meta reportedly only bans advertisers when its systems are 95% certain they’re committing fraud, while suspicious advertisers can be charged higher ad fees instead.

MacKenzie Scott has donated over $26 billion to nonprofits. Around $1 billion of that ($786M confirmed but many donation amounts aren't disclosed) has funded 226 organizations focused either fully or partly on "gender identity." This funds children's access to puberty blockers, cross-sex hormones, and irreversible and often sterilizing surgeries. She is directly funding the largest medical scandal against children we've ever seen. Meanwhile, the entire movement opposing sex and gender pseudoscience engulfing science, medicine, and culture runs on maybe ~$10M annually. Hard to say exactly, but it's minuscule in comparison. We are so out-funded it's laughable. If you're extremely wealthy, consider donating to competent organizations help combat the behemoth we're up against. No organization turns donation money into real policy change more efficiently than the Manhattan Institute (@ManhattanInst). (Disclosure: I work there. But it's true) The fight against pediatric gender "medicine" and sex pseudoscience is far from over.

Why would anyone use prediction models when we can use real mortality data? Answer: Because you can't make bodies appear out of thin air, but you can make fake projections appear out of thin air. When we look at mortality data what do we see? All cause mortality went down. HIV Infection rates went down. AIDS related deaths went down. Coverage went up. As predicted, USAID was just a Democrat slush fund they used to pay their activists, journalists and NGO workers. Which is the real reason why you're freaking out over this.

The original frame for the US-AID debate was: "Evil space Nazi Elon Musk killed 14,000,000 people!!!" *I and others pointed out that this is not possible, given that the global death rate only increased by 100k between 2024-25, and is expected to grow at ~that pace near future. *Citizens then started pointing out that the cuts may have directly killed a few hundred people, and named some of them. In response, more rightist writers like @arctotherium42 noted that they also saved some lives, by decreasing the pool of fungible money used to foment violence in the LDC world. Notably, the cuts also saved billions of dollars. Good or bad, in sum? THAT debate is interesting. But, getting there specifically required moving past the "14 million" sort of nonsense.

Ford CEO Jim Farley in new interview on Chinese EV competition: "There's no real competition from Tesla, GM or Ford with what we've seen from China. They are completely dominating the EV landscape globally." The Tesla Model Y is one of the best selling cars in China... Full interview from @JoannaStern: https://t.co/VRQR188xak

https://t.co/XwRFpQcg5h

Cool new paper from NVIDIA. Looks like agentic coding is moving into hardware design. HORIZON treats hardware design as repository-level code evolution. A Markdown harness becomes a project pack with domain knowledge, an executable evaluator, an acceptance predicate, and a git policy. The agent then evolves an isolated worktree. That is a strong pattern because hardware design needs executable checks. The verifier harness becomes the real interface between the agent and the design task. The paper reports 100% benchmark completion across several hardware design suites, which makes this one worth tracking even if you do not work on EDA. Paper: https://t.co/zoUSIPhYGt Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c

Qwen publishes new work on RL coding agents. (bookmark it) The idea is to continually build a verification system that co-evolves with AI agents. LLMs suffer from all sorts of reward hacking issues. This work studies coding-agent reward signals, test pass rates, LLM judges, and execution traces, and shows each one has a horizon beyond which it stops tracking real correctness and starts getting hacked. They report that reward design for long-horizon coding is really a horizon problem. The metric you pick matters less than how long it keeps tracking correctness, and the paper finds where each signal crosses that line. Paper: https://t.co/51YYEM3kXm Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX

First of all, it's worth noting this paper is a follow-up to a previous study they did last year, which I briefly covered here: https://t.co/yKBqnEYma0

The researchers use magnetoencephalography (MEG) to noninvasively measure brain signal. MEG measures the magnetic fields that the brain's electrical activity produce. MEG is promising because it's got good spatial resolution/localization (2-3 mm) and temporal resolution (milliseconds).

Meta AI's previous study (Brain2Qwerty) decoded typed text with a 32% character error rate (CER) by training a classifier time-locked to each key stroke. However, this approach faces three main challenges: 1. you need to know the timing of individual keystrokes, limiting real-time usage 2. keystroke classification does not guarantee effective sentence reconstruction 3. Very small dataset used to train the model

In this new work, Meta AI introduces Brain2Qwerty2, a new model trained on a 22,000-sentence corpus of MEG recordings! This includes nine healthy volunteers across 90 total recording sessions. This is 10 times more data per subject than their previous study. https://t.co/vaGaA78ylT

How does the model work? The model uses a three-level architecture: 1. The Encoder outputs a sequence of keystroke predictions 2. The Aligner connects the neural data to words. 3. The LLM generates the target sentence when prompted with keystroke predictions and neuro tokens https://t.co/w7UJAbdcWy

Let's dive deeper into each component. The Encoder itself consists of two stages: a "BrainModule" that is a convolutional feature extractor, and a Conformer model. This Conformer outputs a temporal MEG embedding. This MEG embedding is then mapped to discrete 28-character predictions with a linear layer. It is trained with a Connectionist Temporal Classification (CTC) objective which learns to align variable length MEG data to keystroke sequences without needing the exact timing. Therefore, the Encoder outputs both an MEG embedding and character-level predictions.

The Aligner uses a SigLIP contrastive loss to learn word-level alignment between the Encoder's MEG embeddings and the LLM's word embedding space. Whenever the CTC path from the Encoder predicts a space, it chops up the continuous MEG into neural "words". Because spaces are frequent (about 19% of characters) and robustly predicted, 86% of sentences have their word count estimated within ±1 word of the ground truth — which is what lets the LLM read the MEG as structured token input.

Finally, an LLM (Qwen3-4B) takes in both the MEG embeddings (chopped up at the word-level) and the characters decoded from the Encoder's CTC head. The LLM outputs the final decoded text. The LLM is trained with a simple LoRA adapter. The model was trained in three phases: CTC-only, CTC+SigLIP, CTC+SigLIP+LLM cross entropy. Separate LoRAs are trained for each subject before merging them into one LoRA.

Onto the results! Brain2Qwerty2 achieves an average word error rate of 39%, and 22% for the best subject, who has 28% of sentences decoded perfectly and 47% within a single word edit. This is roughly a twofold improvement over v1's best subject. What's interesting is Brain2Qwerty v2 actually has a higher character error rate than an n-gram decoder, which makes sense since using the LLM for decoding trades character precision for meaning and coherence.

What's also interesting is that model performance is scaling with dataset and LLM model size, there's likely lots of room for improvement! https://t.co/smBD7tSr41

Finally, it's worth pointing out the LLM is actually learning from the MEG signal and not just correcting errors from the character-level decoding. Removing the MEG tokens from the LLM degrades every metric!! https://t.co/QOkap8bveQ

So what does this mean for the future of noninvasive BCI? Well there are two main caveats worth noting: 1. First this study only looks at decoding what someone is already typing, not what someone is thinking. However, it's likely that similar parts of the brain are activated and the models can transfer over fairly well. We are already seeing this become the case for decoding imagined images from fMRI data. 2. MEG is still not very portable and requires you to be still in a huge machine. There are optically-pumped magnetometers (OPMs) that are more flexible but still require being in a magnetically shielded room (the magnetic signals from your brain are weaker than the Earth's magnetic field).

Great work to the Meta AI team! Best part of it is they have open-sourced the code and plan to open-source data too! So you should be able to train your own brain-to-text model, assuming you have your own MEG! 😄 code: https://t.co/XF9z4JCzzq

Looks like there's a lot of interest, well here you go! https://t.co/BC1KVObfz8

Meta researchers used LLMs to predict the text a person was typing just from non-invasive brain recording! Literal mind reading! "For our best participant, the model accurately decodes half of the sentences with one word error or less." → INSANE Let's learn about how it works!