This is one of those posts that says something obvious but I think is important/foundational, especially if you're thinking about the future. splittinginfinity.substack.com/p/breakthrou...

It's sunday post papers. I think this one might have implications for LLM inference economics (for small/open source models) arxiv.org/abs/2507.07101

I do increasingly think that open-source ai is of fundamental importance to human liberty, and it's worrying how under-emphasized it is outside of, like, literally just AI2

I can't fully recommend Pluribus. The concept was good but the pacing was *so* slow after the first episode.

Claim that Hyperscalers are spending a bunch of their free cash flow on AI and could run out. xcancel.com/Ross__Hendri...

Dynamic context makes more sense to me than long context. Just put all the tools you need for the immediate task in a medium-sized context window. Grep for new context as needed. I don't attend to my whole life history when I'm doing math problems, why should AI's?

As we enter into 2026: here are a few tips I've developed for posting about AI on Bluesky!

The AGI-2027 authors have moved their timelines back even more it seems www.reddit.com/r/slatestarc...

Squeezing in one more linkpost this year! splittinginfinity.substack.com/p/links-30

This makes me wonder if you could do away with the cable altogether. Like have a 5G phased array on each GPU rack that beams to other racks in the datacenter.

Two companies are making AI chip interconnects using radio waves conducted by plastic waveguides. Claim that it is cheaper, more energy-efficient, and more reliable than optical interconnects. spectrum.ieee.org/rf-over-fiber

Poetiq achieves SOTA on the Arc-AGI-2 semi-private eval. They do it with harnesses on frontier models rather than RL or new models. This task-specific optimization/specialization is The Way. poetiq.ai/posts/arcagi...

I'll have to read this bc it goes against my intuitions. I would think larger batches would have lower variance and training would be more stable across diff hyperparameters!

Great linkpost per usual. Sebastian Seung (leading connectomics researcher) is starting a mind uploading company. Cremieux has "a horribly misguided take." and much more! neurobiology.substack.com/p/action-pot...

Related: physical chunks of neural tissue show brain activity. What are the ethical implications? singularityhub.com/2025/12/25/f...

Zach (chip design expert uninvolved with the deal) thinks Groq's chip isn't competitive, nor are it's partnerships particularly lucrative. Rather, he thinks Nvidia acquired it because they are the only company who designed an inference chip using Samsung's process ... www.zach.be/p/why-did-nv...

Apparently you can buy a laser welding system for $3.5K that makes welding a lot easier. Yet another implication of lasers getting dramatically cheaper. www.youtube.com/watch?v=5to8...

Low-quality simulation of a whole mouse cortex on a supercomputer. How long til a crappy human sim starts talking to us? What are we going to do then? Just turn it off? dl.acm.org/doi/10.1145/...

I haven't had a chance to try them, but a few thoughts on GLM-4.7 and MiniMax-2.1 1. Interesting to see "Context Manage" used to get a 12-15 point bump on BrowseComp. Might be the next phase in AI development? Aligns with my hypothesis that RL and specialization are The Way.

Just wow. @pluralistic.net.web.brid.gy came out swinging in this article. He's right though and that's the coolest part.

Chad Jones has a new paper on AI and automation, looks interesting. web.stanford.edu/~chadj/Jones...

So a bunch of benchmarks likely have errors. If we want AI progress to continue, we need more and better benchmarks. www.futurehouse.org/research-ann...

After staring at enough log-plots, I've decided that they can be kind of misleading. Let me explain ... metr.org/blog/2025-03...

Until we get this, I hope everyone knows that you can *retroactively* limit replies and QTs on a post. Go to the three dots and “edit interaction settings.” I often do that if a post is taking off, even in positive ways; you have no moral obligation to let yourself become discourse.

This is really cool, feels like I'm reading a blog post from the near-future.

Every year, I put together a year's review of what's going on in the crazily evolving world of brain chips, #Neuralink and #BCI's. 2025 was huge for BCIs, but we saw many other trends. Here are my top-10 trends of the year and what to watch for in 2026: #neuroskyence #NeuroAI #NeuroTech

We’re closing in on a “Thermodynamics moment” in AI. A framework for architectures that: - Load info from training data - Meta-learn what info to apply to a particular problem - Perform well on every problem with clean data - Are computationally universal What else do you need?

On why AlphaFold may not revolutionize drug discovery. Great piece by @philipcball.bsky.social

breakthroughs of the year www.renaissancephilanthropy.org/frontier2025

the most recent data shows Waymo's self-driving cars perform significantly *better* than human drivers. the most recent data shows Tesla Robotaxis perform significantly *worse* than human drivers, even with a safety monitor in the front seat. sherwood.news/tech/teslas-...

Detailed review of how we made chips 18 years ago. This is nanotechnology, it is very hard, and leading edge is even harder. www.youtube.com/watch?v=zUgy...

Nano Banana accurately generates images from multiple perspectives. Existing software converts multiple views into a 3D model. So you can use Nano Banana to produce 3D prints! www.youtube.com/watch?v=96OA...

China is still substantially poorer than the west, the part of China we "compete" with is much smaller than the USA. Any shortcomings in human capital can be overcome with immigration. This is a choice; we have chosen not to be competitive.

Michael is on bluesky now. I recommend you follow him!

With Devstral 2, Mistral pushes the envelope on two fronts: ① what an agentic coding model can do, with no reasoning! There’s a quality gap to reasoning models, expectedly. The positive: it is cheaper; potentially even cheaper in practice than indicated in this chart.

Apparently Scott Aaronson figured out how to watermark LLM outputs so that inference providers can check if a piece of text came from their LLM while being undetectable to others. But OpenAI didn't deploy it. Google does something like this though. scottaaronson.blog?p=9333

"In 2023, Nigeria had a million more births than *the whole of Europe*." Humanity's future is quite literally in Sub-Saharan Africa. (1/3)

Great read every year: medium.com/@tomwhitwell...

Here's a great review of what we saw in AI this year, from @gleech.org

It was never conceived this way, but most of the idealized consequences of "solarpunk" can be a direct consequence of this line going down. Radically decentralized greened desert communities can be delivered directly by market forces. caseyhandmer.wordpress.com/2025/12/08/e...

New paper! elliotlipnowski.com/wp-content/u...

In 2023, papers denying motherhood’s role in the gender wage gap went viral. I called out their flawed assumptions then, and newer results have proven me right. www.writingruxandrabio.com/p/of-course-...

It would be neat to see the feed below a post branch into two timelines: one where you liked the post, and one where you didn't. Makes it very obvious what kind of cage you're building for yourself.

New post: The Eldest Millennials Had the Same Fertility as the Youngest Baby Boomers. I did not have a strong take digging into 'fertility crisis' debates, but was genuinely surprised as how different the data is from the panic of that Discourse. mikekonczal.substack.com/p/the-eldest...

Dwarkesh is bearish about AGI in the short term. Bullish long term. www.dwarkesh.com/p/thoughts-o...

on balance, i think i prefer such power to be equalized and open-sourced than held by a few API providers even if that means opening pandora's box

This is exactly what you'd do if you were trying to sell everyone a laptop with AI on it. With 128x compression English Wikipedia fits in 0.2 GB.

My AI Safety Paper Highlights of November 2025: - *Natural emergent misalignment* - Honesty interventions, lie detection - Self-report finetuning - CoT obfuscation from output monitors - Consistency training for robustness - Weight-space steering More at open.substack.com/pub/aisafety...