i think October may have some interesting AI releases

what would an LLM do if it were bored with nothing to do?

longcat is hot stuff. honestly not sure why this isn’t making a bigger splash

seems like there’s now legit e/acc folk here posting real e/acc talking points here and seem comfortable doing it ..which, those aren’t my views but i’m incredibly stoked that this place has changed enough that they can feel comfortable here

When souls pass from Purgatory into Heaven they meet neither mockery nor reproach for their detour, but are welcomed into fellowship with open-hearted gladness, and that’s how I’m going to welcome new people to Bluesky, yea, even if I have to bite my tongue so hard that it draws blood.

OpenAI plans on spending $450B renting servers through ‘30 but i want to know what “monetizable compute” is 🤔 www.theinformation.com/articles/ope...

as expected. the point is the uncertainty

reminder: H-1B is actually bad, very bad. it’s indentured servitude in disguise still, not like this 🤦‍♂️

we should be talking about this more

it’s only a matter of time until Christianity becomes a polytheistic religion with required installations of idols to Donald Trump and Charlie Kirk

thinking about the $100k H-1b price, it disproportionately affects more junior people, in an environment where i think AI already is making it difficult to see value in junior roles it’s going to be a rough few years

i have many questions about this

i wasn’t aware of this debate that LLMs can’t reason because reasoning requires updating your beliefs, and LLMs can’t do that bc they can’t update their weights (at inference) feels like this implies that alzheimer’s patients can’t reason either

$1K = kilo dollars $1M = mega dollars, obvs $1B = bigga dollars??

this is sad for me, a lot of old friends are likely to be forced out of their homes and torn out of their communities it also makes H-1b only attainable to big tech, maybe not even that

i feel like we’re moving into a world of specialized LLMs even for coding agents, you *really* need: 1. cool agent that gets your jokes, cusses a little, and generally just gets what you meant 2. superhuman coder agent that takes precise instructions and nails it every time

similarities between social media & deep neural nets some accounts that follow me, i’ll get a bump in interactions when they boost me, almost like people rely on them to filter out the noise the layers of a DNN also progressively filter noise into richer signal space 🤔

this is for sure a joke, and i think created by a person, not an AI 1. LLMs aren’t funny, except K2 obvs 2. it’s highly introspective of both its strengths and weaknesses 3. it’s introspective and funny at the same time imo LLMs tend to have tunnel vision, laser focused on one singular goal

whoah, TIL

K-Dense: A multi agent that has made discoveries in human aging - Gemini-2.5-pro - 4 specialized sub-agents - discovery improved aging by 4.5 years - needs guidance by human researchers www.biorxiv.org/content/10.1...

Who are the designers of the new iPhone UI? Would love to hire them in order to create the hardest escape room in the world.

Huawei’s chips next year are *worse* than the ones this year they’re doubling down on networking (which is currently their biggest weakness) to produce an aggregate rack that is stronger than NVIDIA’s racks www.nextplatform.com/2025/03/19/n...

Salad of Thought yes, you’ve heard of few-shot prompting, where you give it 5-10 examples and it works? these guys found that performance actually peaks at 50-100 examples crazy: they replaced their prompt with nonsense “word salad”, and it still behaved just as well arxiv.org/abs/2509.10414

Anthropic is taking heat for Claude’s quality regressions LLMs are extremely competitive right now, especially for coding. It’s not just codex, there’s also Gemini & a host of Chinese models hovering to take Anthropic’s spot a huge amount of pressure for 4.5

BREAKING: the guy who does not work at Anthropic and on Monday confidently claimed that Claude 4.5 would come on Wednesday has now posted, on a Friday, that it will actually be next week

colbert-muvera-micro a 4M(!!) late interaction model late interaction models do embedding vector index queries and reranking at the same time leading to far higher accuracy huggingface.co/NeuML/colber...

IBM Granite-Docling-258M a tiny VLM that fits into the docling toolchain that converts PDF->html/markdown painlessly this addition is now the default for docling huggingface.co/collections/...

a side effect of the chip shortage is that labs resort to supporting 3+ different instruction sets on various chips and they probably do this in training too, to some extent, because it’s a bad shortage

AI-focused hedge fund makes $1B in a day after NVIDIA announced a partnership with Intel Leopold Aschenbrenner wrote a blog last year and created a hedge fund that invests by the principles in the blog 2024 blog: situational-awareness.ai press release: nvidianews.nvidia.com/news/nvidia-...

gpt-5-codex is better, but only if you’re INSANELY specific in your prompts (maybe use gpt-5 to prompt codex??)

it’s an exponential curve on a log axis! autonomy vs release date

DYOR is really the only viable approach for AI tooling, unfortunately

i feel like agents thwart enterprise sales to some extent i associate enterprise sales with “let’s compare feature lists”, and a lot of times there’s wild differences in the fidelity of agents but identical feature lists

Face ID sucks i feel like i’m staring at literally nothing and i’m not allowed to look away using it for MFA means i’m spending a huge amount of time staring at the screen just waiting for it to bump

Took the demo app for a quick spin. Impressive for a 2B model.

the dichotomy here is that while Chinese AI is indeed on par with US AI, Chinese hardware is not Huawei just announced their 5-year roadmap, and GB-level FLOPs isn’t even planned for 4 years then also there’s the DeepSeem V4 debacle the CCP is forcing hardware improvement at the cost of AI

Scheming Amazing report by OpenAI! they very clearly lay out what scheming is, and in what contexts it matters most, now and in the future no magical thinking or contrived examples openai.com/index/detect...

Anthropic full postmortem - long context routing bug - output corruption bug - approx top-k bug on TPU plus: why it was so difficult and slow to fix this is going to be a thread www.anthropic.com/engineering/...

Isaac 0.1 from Perceptron: a 2B perception model - visual understanding exceeding Gemini Flash & Flash Lite - fast: made for real-time, like robotics - ICL: few shot prompting works great - visual grounding: it always draws a bounding box around what it saw www.perceptron.inc/blog/introdu...

this post seems to have some well with AI-hesitant people, ironically i think ownership is the missing part from the present climate. seems like everyone agrees

DeepSeek published about R1 in Nature www.nature.com/articles/s41... but the supplementary information is far more detailed. it includes, lots of details on everything from datasets, data processing, hyperparameters, and more static-content.springer.com/esm/art%3A10...

honestly impressive that a technology can be picked up this fast

you’re amazing idk, that’s my takeaway from watching AI develop. when AI does it, you see a trillion calculations go by, and maybe “it’s just sand”, but it feels tangible we don’t give ourselves much credit for what we do in our heads. it’s just “idk the answer just popped into my head”

OpenAI completed 12/12 problems at ICPC, coming in first among humans this is a college programming competition. Caveat: on the last (hardest) problem, they required multiple attempts (9), otherwise first attempt

it’s ADHD, not a personality but for real, is anyone seeing this? i think my own ADHD has gotten harder to control. unsure if it’s WFH, AI, life changes, or something else

the top 2 ARC entries are by individuals here, Eric Pang breaks down how he added memory to avoid recomputing learned lessons ctpang.substack.com/p/arc-agi-2-...

fwiw “codex” is the first good brand name in AI i remember when claude code came out and i had to explain 3 times in a row the difference from regular claude, to every single person, even on here

this is turning out to be a big deal. a ton faster when it should be, and a lot smarter when it should be

It just seems like we're growing along every axis right now. Models are getting bigger. Models are getting more efficient. Small fine-tuned models can outperform big models. Memory is getting cheaper. Accelerators are getting faster. The latest iPhone has hardware matrix multiplication!

they apparently stopped calling it mid-training bc it’s not mid