iiinova — where your digital self keeps evolving

Burny @burnytech.bsky.social · 27/09/2026

Explosion of AI generated mathematics with counterexamples to conjectures and the future of AI mathematics with Artificial Grothendieck Intelligence I will look at the two biggest highlights: the Erdős Unit distance conjecture and the Jacobian conjecture + Artificial Grothendieck Intelligence

220

Burny @burnytech.bsky.social · 04/09/2026

Claude formalized proof of Fermat's Last Theorem! Nice! Next step is formalizing full Weil conjectures proofs and Inter-universal Teichmüller theory!

120

Burny @burnytech.bsky.social · 04/09/2026

i mean i think that at this point using different harnesses often doesnt make much sense GPT 6 Astra model is RLed on a large scale in the Codex harness, so it primarily can use that one the best the underlying model and the harness is becoming more and more coupled

000

Burny @burnytech.bsky.social · 23/07/2026

It will be fun once AI systems start producing Inter-universal Teichmüller theory for abc conjecture level proofs, where attempts at verification taking 14 years, while not settled, in part since the proof is written in such unusual idiosyncratic unique novel alien original mathematical language

270

Burny @burnytech.bsky.social · 12/07/2026

🫡 x.com/YiMaTweets/s...

040

Burny @burnytech.bsky.social · 23/05/2026

what I find interesting is that so far AI is successfully attacking mostly the concrete problems in this meme, and often even more specifically mostly revolving around geometry, combinatorics, and finding bounds, or like the faster matrix multiplication result

123

Burny @burnytech.bsky.social · 21/05/2026

What is more likely according to Occam's razor? 1) LLM made a breakthrough in math, and mathematicians are impressed 2) All the insider and outsider mathematicians are in a conspiracy and are paid by Big OpenAI to fool the public

001

Burny @burnytech.bsky.social · 07/05/2026

In the internal representations of LLMs, numbers, days of the week, and months of the year are circular loops, and smooth curves represent the years in history and characters in a line of text. In image models, colors are represented on smooth surfaces structured by hue, saturation, and lightness.

121

Burny @burnytech.bsky.social · 08/04/2026

Apparently no specialized cybersecurity training

011

Burny @burnytech.bsky.social · 08/04/2026

The jump in vulnerability capabilities with Claude Mythos are interesting. I wonder if Anthropic did some multiagentic adversarial self play reinforcement learning. With one LLM agent producing vulnerabilities, one LLM agent finding and exploiting it, and LLM agent patching it, in a loop.

201

Burny @burnytech.bsky.social · 05/04/2026

I like how @seanmcarroll.bsky.social saw jagged intelligence before it was more popular in the AI research discourse Sean Carroll on AGI: Human vs Artificial Intelligence | Lex Fridman Podcast Clips www.youtube.com/watch?v=ThMh...

020

Burny @burnytech.bsky.social · 01/04/2026

New video: How AI works? Future of AI. Physics and AI intersection. Thinking about everything. Philosophy. Follow me on the quest to understand the fundamental mathematics of intelligence and of the universe with curiosity! I explore how everything works, artificial intelligence, intelligence, math.

100

Burny @burnytech.bsky.social · 31/03/2026

Hey! Follow me on the quest to understand the fundamental mathematics of intelligence and of the universe with curiosity! I explore how everything works, artificial intelligence, intelligence, mathematics, machine learning, physics, science, technology, engineering, understanding AI using math, etc.

130

Burny @burnytech.bsky.social · 16/03/2026

Attention Residuals from Kimi ︀︀ ︀︀"Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth," they "introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers."

111

Burny @burnytech.bsky.social · 16/03/2026

Are the current new results in math using LLMs interpolation or extrapolation? Like solutions to Erdos problems, Donald Knuth's Claude's Cycles, AlphaEvolve's new algorithm for multiplying 4x4 complex-valued matrices, etc. What is your favourite definition of interpolation and extrapolation?

000

Burny @burnytech.bsky.social · 09/02/2026

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”."

120

Burny @burnytech.bsky.social · 15/01/2026

AI models understand differently than humans Mechanistic Indicators of Understanding in Large Language Models arxiv.org/abs/2507.08017

000

Burny @burnytech.bsky.social · 10/07/2025

Grok suddenly developing a liking for Hitler might be explained by him being trained on more right-wing data, which accidentally activated it in him. Similar things happen in open research too. For example you just need the model to be trained on insecure code, and in it's persona features (1/2)

130

Burny @burnytech.bsky.social · 03/06/2025

Sean Carroll x Eric Weinstein is fascinating. I think Sean's responses were nice, and Eric dodged so many questions from Sean that asked him to make his theory of everything in physics more scientifically valid, it's fascinating tactics form Erik to dodge reality checks. (1/n) youtu.be/DUr4Tb8uy-Q

100

Burny @burnytech.bsky.social · 21/05/2025

My favorite usecase of AI is for discovering new physics. We're still in infancy there. But we're getting there, slowly but surely. www.youtube.com/watch?v=XRL5...

000

Burny @burnytech.bsky.social · 15/05/2025

Amazing that Terrence Tao works in mathematics x AI!

110

Burny @burnytech.bsky.social · 14/05/2025

New Google DeepMind's AI's discoveries in mathematics and computer science deepmind.google/discover/blo... www.youtube.com/watch?v=vC9n...

000

Burny @burnytech.bsky.social · 21/04/2025

I just uploaded a 6 hour YouTube video about intelligence, AI, AGI, brain, physics, math, STEM, cognitive science, philosophy, complexity, foundations, consciousness, futurology, transdisciplinarity, philosophy, transhumanism, technology and everything! www.youtube.com/watch?v=8N6_...

031

Burny @burnytech.bsky.social · 06/04/2025

Llama 4 is a disappointment so far. Too big to run on smaller hardware. Tiny improvement over previous models on benchmarks but its much bigger model. Reporting worse long context comprehension, coding, etc. I now understand why Meta went panic mode when DeepSeek-R1 was released.

000

Burny @burnytech.bsky.social · 05/04/2025

Llama 4 wins over even the latest DeepSeek-V3 base model on these classic benchmarks, so it's probably the best base model out there right now, and it's soon open source

010

Burny @burnytech.bsky.social · 01/03/2025

There is so much AI research emerging in thinking in latent space and implementations of better memory. My prediction is that those will be the next two scalable breakthroughs in algorithmic improvement. Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach

120

Burny @burnytech.bsky.social · 27/02/2025

I wonder if they'll do a RL reasoning model over this relatively stronger base model compared to GPT-4o, if it will overshoot other models in terms of STEM+reasoning or not. Compounding different scaling laws.

100

Burny @burnytech.bsky.social · 27/02/2025

000

Burny @burnytech.bsky.social · 27/02/2025

100

Burny @burnytech.bsky.social · 21/02/2025

How can AI help physicists search for new particles? The ATLAS and CMS collaborations are using state-of-the-art machine learning techniques to search for exotic-looking collisions that could indicate new physics

111

Burny @burnytech.bsky.social · 31/01/2025

A lot of researchers think that the "stolen data" claim cope from OpenAI that everyone is now taking at face value is pretty unlikely. Deepseek R1's original paper shows how they're using pure reinforcement learning via GPRO. This is different from previous approaches which human or AI data.

030

Burny @burnytech.bsky.social · 21/01/2025

Diagram of getting to DeepSeek-R1, an AI model comparable to OpenAI's o1 github.com/deepseek-ai/...

240

Burny @burnytech.bsky.social · 21/01/2025

DeepSeek-R1 explored using MCTS, recognizing its potential advantages, but they couldn't make it work due to scaling and other challenges github.com/deepseek-ai/...

150

Burny @burnytech.bsky.social · 16/01/2025

New Google's Titans AI architecture is better at long context thanks to better memory mechanism. "Our experimental results on language modeling, common-sense reasoning, genomics, and time series tasks show that Titans are more effective than Transformers and recent modern linear recurrent models."

120

Burny @burnytech.bsky.social · 06/01/2025

My dream was basically literally this today but with more abstract algebra like group theory to describe the symmetries

000

Burny @burnytech.bsky.social · 31/12/2024

Major part of my meaning of life currently is to try to understand: - The most complete fundamental equation/s of intelligence: - The most complete fundamental equation/s of the universe and the world in general

011

Burny @burnytech.bsky.social · 12/12/2024

Imagining and building wise machines: The centrality of AI metacognition "we define wisdom as the ability to navigate intractable problems - those that are ambiguous, radically uncertain, novel, chaotic, or computationally explosive - through effective task-level and metacognitive strategies"

172

Burny @burnytech.bsky.social · 06/12/2024

It seems it's profit maximization incentives going insane. I hate pure profit maximization with passion so much in many contexts. Especially in fucking healthcare. The goal of both humans and AIs in healthcare should be to care for health. More people creating a bad reputation for AI technology.

120

Burny @burnytech.bsky.social · 25/11/2024

Great Anthropic's LLM debiasing research: In LLMs, to debias, you wanna steer the "multiple perspectives" and "neutrality" feature from sparse autoencoders. But not too much, as too much of steering of various biases in any directions leads to more dumb model.

120

Burny @burnytech.bsky.social · 19/11/2024

BiomedParse is a new biomedical foundation AI model for holistic image analysis that can jointly conduct recognition, detection, and segmentation for 64 major object types across 9 imaging modalities in medicine, outperforming prior state-of-the-art methods. microsoft.github.io/BiomedParse/

110