Burny

GPT-5.2 derives a new result in theoretical physics | Discussion

I think about this often

When Models Manipulate Manifolds: The Geometry of a Counting Task We find geometric structure underlying the mechanisms of a fundamental language model behavior.

transformer-circuits.pub/2025/linebre...

My favorite reverse engineering how LLMs work internally research from Anthropic

They "find learned representations of position and find dual interpretations:" they "can understand them as a family of discrete features or as a one-dimensional “feature manifold”/“multidimensional feature”."

A smooth galaxy from the Galaxy Zoo: Hubble project, classified by 51 volunteers.

A smooth galaxy, observed with the Hubble Space Telescope in the COSMOS survey.

It is at redshift 0.23 (lookback time 2.84 billion years) with coordinates (150.29382, 1.65700).

51 volunteers classified this galaxy in Galaxy Zoo: Hubble.

The only way we dont have AI is if you invented time travel and mercilessly slaughtered everyone involved with matrix multiplication in history, every time its discovered, forever. Good luck banning math.

Thats not a real thing that you can do. You can learn how to use it effectively and teach your students how to use it safely which you should. There *is no stopping it* because its already proven its a viable product. The question is how do we live with it, not do we live with it.

In a rush, I approved a plan for Claude to train on the eval dataset 🤦

M78: Reflecting Blue in a Sea of Red apod.nasa.gov/apod/ap26012...
In the Orion Molecular Cloud complex, several bright blue nebulas are particularly apparent. Pictured here in the center are 2 of the most prominent reflection nebulas - dust clouds lit by reflecting light of bright embedded stars

Holy moly this chart: Cumulative US measles cases

Large-Scale Identification of Novel Protein Biomarkers and Therapeutic Targets in Heart and Brain Disease https://www.medrxiv.org/content/10.64898/2026.01.26.26344874v1

ESA/Hubble photo of a distant stellar birthplace, region of the N159 star-forming complex in the Large Magellanic Cloud, approximately 160 000 light-years away. Photo credit: ESA/Hubble & NASA, R. Indebetouw

Link for more information: esahubble.org/images/potw2...

waiting for some experiments to run, so a quick thread about base models and pretraining contamination, with some weird & interesting base model generations i've collected over time.

or, why do open source models claim to be claude or chatgpt?

Me in 2005: Moore's Law means that computers will get super fast over a very short period of time!

Me in 2026: my computer freezes up if I leave my browser open for too long

The Pleiades star cluster (M45) captured over 25 hours. The blue glow is starlight reflecting off interstellar dust. The surrounding brown structures are the Integrated Flux Nebula, incredibly faint clouds lit by the glow of our entire galaxy.
#astrophotography #astronomy #space #m45

GitHub - antfu/skills: Anthony Fu's curated collection of agent skills. Anthony Fu's curated collection of agent skills. Contribute to antfu/skills development by creating an account on GitHub.

late to the party. I have finally been convinced by multiple awesome developers to give agents another try.

this is my first premature contribution:
github.com/antfu/skills

GitHub - google/A2UI Contribute to google/A2UI development by creating an account on GitHub.

A2UI: Agent-to-User Interface

3-panel comic. (1) [Three small arthropods on ocean floor.] ARTHROPOD 1: Now that we’re multicellular, what are your plans? I’m gonna evolve little legs and swim around with them! ARTHROPOD 2: I’m gonna evolve sharp pincers and use them to crunch stuff! ARTHROPOD 3: I’m gonna evolve glands to make string from my butt and use it to construct elaborate geometric nets hundreds of times my size to catch other animals. (2) [Silence] (3) ARTHROPOD 1: *Dude.* ARTHROPOD 2: Can you *please* just be normal about this? ARTHROPOD 3: *What??!*

Early Arthropods

xkcd.com/3199/

Subthreshold Kir and Ih currents modulate excitability of Layer 1 VIP interneurons in the medial prefrontal cortex https://www.biorxiv.org/content/10.64898/2026.01.28.702118v1

I got long email with research questions, full with long equations written in LaTeX. So instead of compiling this with LaTeX, I dumped it into Gemini, asking it to render the equations for better readability.

Gemini did it ... and started to answer the questions, which it was not asked to do 😂.

Is the Policy Gradient a Gradient? The policy gradient theorem describes the gradient of the expected discounted return with respect to an agent's policy parameters. However, most policy gradient methods drop the discount factor from t...

Writing the policy gradient lecture, which gives a great opportunity to discuss the sneaky missing discount factor that makes almost every paper wrong:
arxiv.org/abs/1906.07073

Search for τ⁻→μ⁻μ⁻μ⁺ decays at the LHCb experiment with Run 2 data arxiv.org/abs/2601.20785 - An upper limit of 1.9(2.3)×10⁻⁸ is set at the 90% (95%) confidence level on the branching fraction of the decay.

I view AI image generation as a sort of new type of computer graphics pipeline. We constantly use computers to render graphics so probably these systems will find a lot of uses, especially as they improve and we distill the algorithms

But we had to learn how to use photography. Learn the mechanics.

Converging evidence of positive selection at height-associated loci in Europe https://www.biorxiv.org/content/10.64898/2026.01.27.702172v1

Engagement, user satisfaction, and the amplification of divisive content on social media Abstract. Social media ranking algorithms typically optimize for users’ revealed preferences, i.e. user engagement such as clicks, shares, and likes. Many

This is pretty fascinating: academic.oup.com/pnasnexus/ar...
(1) Twitter's algorithm makes its users feel worse about their outgroup
(2) Users do not prefer the tweets it suggests relative to simple baselines

Canaries in the coal mine. Worth paying attention to.

(And yes, they are both obviously interested in seeing their own products used, but hearing enough from other, independent coders that make me believe them. I wrote more about the shift here: www.oneusefulthing.org/p/management...)

I’ve always liked programming. But I’ve been hearing similar reports from engineers who I respect, and I’m seeing this shift in my own work. IDK what the future holds, but programming as we knew it has already changed.

Text Shot: While "thinking" models are powerful, they have historically been siloed — great at math, but poor at browsing the web or running code. Qwen3-Max-Thinking bridges this gap by effectively integrating "thinking and non-thinking modes". The model features adaptive tool-use capabilities, meaning it autonomously selects the right tool for the job without manual user prompting. It can seamlessly toggle between: Web Search & Extraction: For real-time factual queries. Memory: To store and recall user-specific context. Code Interpreter: To write and execute Python snippets for computational tasks. In "Thinking Mode," the model supports these tools simultaneously. This capability is critical for enterprise applications where a model might need to verify a fact (Search), calculate a projection (Code Interpreter), and then reason about the strategic implication (Thinking) all in one turn. Empirically, the team notes that this combination "effectively mitigates hallucinations," as…

Qwen3-max-thinking-beats-gemini-3-pro-and-gpt-5-2-on-humanitys-last-exam venturebeat.com/technology/qwe… #AI #Qwen

Amazing

Semantic axes in the brain support analogical representations https://www.biorxiv.org/content/10.64898/2026.01.28.702241v1