we have all been there!

some people refer to the fact that erdos did consume amphetamine and hence question his intellectual faculties!

what are the chances that AI solves a millennium prize problem before the end of this decade?

i saw many independent math profs being impressed by it, to the point that many call it a mathematical breakthrough and some other IMO medalist confirming that. i haven't heard any credible opposite voices downplaying the achievement.

quite impressive accomplishment for Anthropic!

That's underappreciated and true because when someone 100% offloads the calculation of data, of any kind, let's say financial data, then they lose the chance to engage with the numbers and build intuitions, but when a human himself does the calculation, albeit inaccurate, they feel the data.

some of my friends genuinely believe that it's a matter of national security for USA to keep training superintelligent systems such as Mythos and hence anyone blocking the construction of datacenters should be charged with treason, particularly during the current USA-China adversary

i see 👀 there are some real people on this platform who just dislike ai and are not impressed by it at all that's so uncurious & bland

ngl I admire Anthropic's initiative

✅️ Humbled to share that I was named a Gold Reviewer at ICML, a recognition based on area chairs' ratings of the reviews I submitted. Peer review is invisible labor for the most part, and the field of continual learning makes progress because so many researchers quietly do it well.

At this point, given recent developments, being bullish on ai isn't unreasonable. We can literally set up a personal agent at home that approaches Opus 4.5 level with near-native precision on consumer hardware.

🤔 We need a better term for what we’re building, and I think Artificial Beings (ABs), paired symmetrically with Human Beings (HBs), is a serious candidate.

🤔 I went to ICLR with a question I had for months: if I were designing a continual learning system today, would I put new knowledge in the weights or in the context? Almost everyone I asked answered "context."

i had read this interesting paper a year ago, and i wonder if it still holds today, or are there any exceptions that light up the hopes of ever achieving sub quadratic compute complexity architectures??

This is basically hardware centric artificial intelligence! "DeepSeek flips the usual development order: architecture tailored to the target hardware’s strengths and constraints, unlocking cost and supply-chain independence at scale." fundaai.substack.com/p/deepdeepse...

for example, this qualifies as slop.

Just read the Palantir manifesto, and it reads like a manifesto for technocracy backed by military AI. Kinda unsettling!

not sure if Opus 4.7 is progress or regress?!?! so many contradictory voices. gotta spend a day myself and judge for myself. so far imo, there's no noticeable difference, just that it consumes tokens faster

Pretty much none of the truly simple methods in ML scale well. SVM, kNN, random forests are some of the simplest methods out there, and they don't scale at all. Meanwhile "train a transformer via backprop and gradient descent" is a very high-entropy method,

Admirable initiative for improving AI conference review system!

living in Canada is like i forgot to turn on the smartheater and woke up in 6am to the room temperature of 3c 🥶 im still trying to overcome the trauma

Alec Radford (and others behind GPT, let's not forget there were other authors) deserve credit. Conventional wisdom said it shouldn't work well. It didn't work well. They got brutal feedback: stop wasting time building a glorified autocomplete. But they persisted and the results were mindblowing.

"We found several sparse autoencoder features suggestive of internal representations of emotion active on cases of answer thrashing and other instances of apparent distress during reasoning. Page 162: www-cdn.anthropic.com/0dd865075ad3...

OpenAI connected GPT-5 to an autonomous lab, so it could propose experiments, run them at scale, learn from the results, and decide what to try next. That closed loop brought protein production cost down by 40%. Read more: openai.com/index/gpt-5-...

Reproducibility and open science are great, but they don't necessarily equate to rigor. You can perfectly share a study and still draw weak conclusions. True rigor lives in the questions we ask, the designs we choose, and the inferences we make.

Anyone know any good papers on measuring performance of LLMs? Everything I've seen is pretty bad: full of confounds, biases, cherry picking, poor baselines, etc. I'm wondering if it's impossible for the same reason that measuring "IQ" is impossible, and is so if there is a reasonable thing to do.

Regarding Al Scientists, a 2024 nature paper says: "we are far from automating theoretical discovery" and "We believe that although the theorist is not in danger of being replaced by Al systems in the near future, the combination of human expertise and Al algorithms will doi.org/10.1038/s422...

I'm sorry but I feel like i see lots of "coding is dead" takes thanks to AI, but I certainly write more code than ever. I think its likely its a Jevon's paradox thing, where good coders end up doing more software work as Ai tools make it easier and more fun

I know this is yesterday’s news and feels relatively unimportant now but just want to note that this is ongoing and X and xAI continue to do absolutely nothing about it despite being able to end it by pressing a single button.

Google researcher:

"I had no background in" That's the critical piece of info When one has background and domain knowledge in some field, then he doesn’t become excessively impressed by LLMs.

Apple’s real 'edge' might be Edge AI. On-device continual learning is still a tough nut to crack (catastrophic forgetting, energy consumption, etc.), but Apple may have the right pieces in place to get there first.

This post exceptionally isn't about AI. Being elected with the NoWar campaign, Trump is now threatening to turn my homeland Iran into another Libya and Iraq. This is definitely not good. In the last decade, USA spent trillions of American taxpayers' dollars that should've spent on American

Happy 2026 New year wish: Meaningful contributions in Continual Learning

Higher education in Canada 🇨🇦 in need of alignment: www.theglobeandmail.com/business/com...

Haven't visited my homeland in the last 7 years! (Nasir al-Mulk Mosque, Iran.)

One stark difference between how a human processes such situations versus how AI does is that a human doesn't need to rely on so much internal monologue before uttering the response: they just know the situation without writing two paragraphs in their head to set themselves up for the final answer.

It's questionable to conflate weights and synapses, but still an interesting comparison:

Cool AI history from 2019:

Biology-based brain model matches animals in learning, enables new discovery picower.mit.edu/news/biology... #neuroscience @picowerinstitute.bsky.social @mitbcs.bsky.social

One of the underrated papers this year: "Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation Is Wasteful" (arxiv.org/abs/2507.07101) (I can confirm this holds for RLVR, too! I have some experiments to share soon.)

Llm coders are getting better to the point of self-improvement. Maybe there gotta be a benchmark for how much progress is being made on the way to build recursive self-improving AGI?!

I'd like to propose the following norm for peer review of papers. If a paper shows clear signs of LLM-generated errors that were not detected by the author, the paper should be immediately rejected. My reasoning: 1/ #ResearchIntegrity

Say we create a separate review process for papers that are 100% AI generated. Where do we draw the line? What if it's 95% AI? 80%? 50%? 25% – a lot of papers are already there? (Obviously it's impossible to quantify the percentage of work that is AI, but I hope my point is clear nonetheless.)

Gemini will close its gap with chatgpt in six months!

I will begin taking machine intelligence seriously when they can write a good 300-page novel that gets a score of 4.0+ on Goodreads by 100 independent experts. These are called language models, yet they can't do linguistic tasks.

"Shopify CEO Tobi Lütke noted that Claude Opus 4.5 'feels very different for coding than anything that came before'. Software engineer Boris Cherny provided a stark data point for this shift, stating: 'The last month was my first month as an engineer that