Kimi K2-Thinking a new leader? moonshotai.github.io/Kimi-K2/thin...

OpenAI has been getting ready to release GPT-5.1 (this from their iOS code) pretty sure i’ve A/B tested it, and it was a big step up, at least for the search-type queries i typically do

looks interesting

@cameron.pfiffer.org what’s the idea here? just seeing if an agent will become anything useful without curation?

Apple accidentally leaked that it’s using a 1.2T model from Google so is that Gemini 3 Flash, Pro or Ultra that’s 1.2 T? fwiw, general speculation says Google wouldn’t give Apple a Pro model www.bloomberg.com/news/article...

🔨 this is not a drill!

why is gemini 3 delayed? the answers may SHOCK you reason 1: the mixture of experts got too opinionated and now they can't agree on anything reason 2: they designed it as a Matryoshka transformer, but they misplaced one of them and they're still scrambling to find it

are we FINALLY getting async cobol?

is the supreme court broadcasting audio live on CNN? is that new? i thought audio was hard to get a hold of

PSA: MCP is fine and not scary at all if you use it to only pull information like if your pulling data from 6 different databases, use MCP, it’s fine Just disable tools that can edit or delete (or do anything that might make you sad if done wrong)

Lily is one of my coworkers. imo she's the archetype of what success looks like for AI in the workforce she's in marketing, and leverages AI in ridiculously powerful ways i like her phrasing — "Ops" is the superpower, not AI. AI is merely the tool www.appliedaiformops.com/p/why-ops-sk...

overheard: “they’re on twitter, instagram, X.. i don’t even know what X is, what is X?”

we didn’t see many US AI releases in October. i had been predicting a lot more best case: AI slowed down worst case: it sped up but nothing’s being released

Anthropic predicting twice the revenue as OpenAI in ‘28, regarding selling AI to businesses I imagine OpenAI’s remaining business revenue will boil off as the Microsoft exclusivity deal ends www.theinformation.com/articles/ant...

Windsurf Codemaps actually this makes a ton of sense — if vibe coding only works on small/non-complex projects, then the answer is to tackle complexity directly Codemaps uses LLMs to create an “index” over your code, a map of where things are cognition.ai/blog/codemaps

Starcloud: GPUs in space This company finally launched their first H100 into high Earth orbit. A solar array for power, uninterrupted by weather or nighttime, and a black plate in the back to radiate heat away into -270°C space starcloudinc.github.io/wp.pdf

Thinking in both text & image leads to new emergent properties researchers SFT’d a small 7B on reasoning traces that use both image and text for reasoning results: - huge 👍 on benchies - emergent properties, like using image or text reasoning at the right time thinkmorph.github.io

Anthropic Model Depreciation Process Anthropic sweetly asked Sonnet about its preferences in how it wanted to be deprecated in addition: - no, still not open weights - preserve weights and keeping it running internally - letting models pursue their interests www.anthropic.com/research/dep...

the fact that napalm death doesn’t have a song called “996” is an indictment of the metal scene

linkedin seems to heavily penalize dormant accounts, so i try to leave a slow steady trickle of clickbait posts there in case i ever need to get big reach on a post

this seems like a huge accelerant to climate and earth science anyone out there more familiar with exactly how this thing is used? ideally specific use cases..

OpenAI: We're buying $743B in GPUs from Mayo Clinic Mayo Clinic: Yes, we're very excited that OpenAI is interested in Geriatric Psychiatric health, it's long past due

I added this to my AGENTS.md file (text in alt) and it seems to work well i had an environment error, spun out a new codex-cli to figure it out, it wrote a lesson. my other codex instances can see it and benefit from it

when i have more time, i want to figure out why anything with flash attention ends up being so f huge probably something @advanced-eschatonics.com knows about want a vLLM docker image? cool, 2KB. oh, with flash attention? that’s 2TB

Consistency Training new GDM research notes that both jailbreaking and sycophancy share a common cause — subtle changes in the prompt cause dramatic changes in output they address it by training on small prompt changes, expecting the same result deepmindsafetyresearch.medium.com/consistency-...

i’ve been thinking about this a lot the last month, finally wrote it up i think it enables truly general agents to operate with oversight, but also brings sanity to the current state of AI security

this is a nightmare for alignment research, but it shouldn’t be this should be a tool for peeking inside LLM thought process

MCP Colors A riff off of the lethal trifecta for addressing prompt injection, this is a simple heuristic to ensure security at runtime red = untrusted content blue = potentially critical actions An agent can't be allowed to do both timkellogg.me/blog/2025/11...

Cache to Cache: let agents communicate in KV cache latent space Instead of con concatenating the text from one agent into another, just concatenate their KV cache directly this is dumb, how do i get it now??? fuvty.github.io/C2C_Project_...

all y’all who think LLMs are too sycophantic haven’t talked to a 5yo

BREAKING: the Microsoft <-> OpenAI alliance

1st package manager result for "wat": pypi: Deep inspection of Python objects npmjs: Community-controlled cheat sheets for every coder. crates.io: Rust parser for the WebAssembly Text format, WAT go: Package wat is a generated protocol buffer package.

happy monday

yesterday i got a A/B test on chatgpt and the alternative was *really* good the thought process was very structured. it gave me both depth and breadth at the same time

Rule of Two: fighting prompt injection @simonwillison.net posted on a new phrasing of the Lethal Trifecta that changes one node to “externally communicate OR change state” in my own work, my version was “OR perform critical actions” simonwillison.net/2025/Nov/2/n...

this is getting ridiculous

it’s plausible that Gemini is ~10T, and GDM has been clear that the next Gemini is 10x the size that’s big enough that it can’t fit onto a single server rack. imagine the crazy problems you face at that scale like, there’s no way it’s serving one request at a time. no chance

what’s the most pointless LLM emergent behavior? a behavior that clearly doesn’t help on benchmarks or make a product “better” maybe eval awareness & alignment faking?? although arguably that helps even more since it effectively fabricates perfect alignment scores

GB300’s cost $3M per rack that’s for an NVL72, i.e. rack of 72 B300 GPUs along with some shared memory boost, some CPUs, etc. so i guess that puts a single B300 at around $40k www.barrons.com/livecoverage...

Training LLMs end to end is hard. But way more people should, and will, be doing it in the future. The @hf.co Research team is excited to share their new e-book that covers the full pipeline: · pre-training, · post-training, · infra. 200+ pages of what worked and what didn’t. ⤵️

going around my house measuring watt usage on various devices and holy wow people who are concerned about AI electricity use: do you drink tea or coffee?

my most autistic behavior is needing consistency i’m moving everything out of my office so we can get work done on the floors and it is stressing me tf out i just moved a bookshelf out and i need to take a tea break to calm down

bf16 halloween might be already ending. according to a bytedance engineer could just have been another flash-attention bug.

i feel like this is a good summary of the current state of AI agents & models

people here & everywhere bitterly disagree whenever AI consciousness comes up imo that’s amazing, very encouraging. it’s a topic that makes us question our very existence we’d have to be truly lulled into senselessness to not bitterly disagree over it

LLMs can report their own experience the most convincing experiment: they isolated the deception control vector and had it talk about its own consciousness it more openly introspected when deception was *suppressed*, and boosting reduced introspection arxiv.org/abs/2510.24797

looked into it, codex-cli switched to Rust because: 1. static binary / easy distribution 2. easier OS-level security controls 3. perf 4. extensibility — they want to introduce wire-level extensions that can be written in a variety of languages

that’s codex-cli and claude code now is Rust taking over AI? what’s the advantage of Rust here?

PSA if you think a news article is dumb, DON’T share it. That just makes the problem worse

holy shit, yes, such a good thread, the entire thing