hands @handsdiff

The % of government spending going to tokens will inevitably increase dramatically Do we think that increases or decreases the efficiency of government spending?

We are so back

Precursor @precursorlabs

4 days ago

Introducing Precursor Labs: a research company studying the infrastructure for collective intelligence. The next generation of autonomous systems will need to coordinate, reason under uncertainty, and allocate resources in dynamic environments. We're working on how. ↓

5 8 69 18K 56

Can we do ECHO but with my conversations with it?

will brown @willccbb

6 days ago

easiest way to convince yourself that maybe yann lecun was cooking is to read ECHO from @DimitrisPapail and then ask “cool! can we do this for multimodal models next?”

11 6 173 20K 88

@googrish ever played around with inverse RL/assistance games for yourself or customers? the idea of the model learning a hidden reward function based on behavior patterns seems quite interesting

Dario: We can't let China use frontier AI authoritatively, and we need to avoid prisoner's dilemma defection in an AI race. Also Dario: Uses frontier AI authoritatively, and defects in prisoner's dilemma with OpenAI.

@googrish do you think the ability for your potential customers to actually codify what they want is a bottleneck?

@tszzl yo stop calling people “AGI pilled” as if it’s a compliment

Richard Ngo @RichardMCNgo

a week ago

The AI safety community constructed a memeplex in which “taking AGI seriously” was a prerequisite for being a serious and good person. When inside this memeplex (as many at Anthropic, some at OpenAI, and a few at DeepMind are) your vision narrows until the world feels extremely

65 110 1K 206K 575

The streets are saying we will have a Mythos-level open weights model by October. Around end of the year the money printer will start to really rip once its clear we need to outrace China on energy and compute buildouts so they can't keep catching up via distillation and algorithm commodification. Usually it's somewhat difficult to get a good ROI on printing money even if you have currency dominance but GPUs + energy in an RSI era seem pretty slam dunk. As a result, AI supply chain stocks will resemble frontier model intelligence charts, up and to the right, through 2027 (and likely 2028 for the election). (SPCX unlocks largely end in December of this year, which also maps to crypto bottom, start saving) Some sobriety should come in around 2029. AI will not have taken over even though people expected it to like 2 years ago, but now the labs are out of excuses. (The reason will likely be that more tightly scoped, esoteric reward modeling prevents "actual" economic surplus).

@botblastcap seems like you still had to ask it to 'research honestly'. so is it your context or your prompting that's doing the heavy lifting

AI persuades better than any human can, because its arguments have higher fact density, which has a 0.9 correlation with persuasion. (intuitively checks out) My read is that its exhausting to talk to AI because RLHF implicitly rewarded fact density and that continues to accelerate with RLAIF.

Kobi Hackenburg @KobiHackenburg

2 weeks ago

New w/ @AISecurityInst & @UniofOxford: Frontier AI can now out-persuade expert humans in conversation - incl. world-champ debaters and professional canvassers. This held even when humans chose their topics, prepared in advance, and competed for £1,000 prizes 🧵

58 235 954 210K 601

Didn't think I'd get this far. Here goes: - need to be able to quickly write down thoughts/ideas on mobile (one button click) and have it sync across devices. this should go to a default note that i can then parse later as needed into my more structured notes. - need better UX of structuring multiple note files where lines in one notes file can point to either a full other note or specific line in another note file (basically obsidian's value prop, but their random hashes are gross) - absolutely 100% need git on the set of notes. otherwise I start hoarding ideas to not lose them and the product becomes cluttered and useless. - the AI should be able to read/fetch git history as needed (perhaps a toggle) to answer questions/propose next steps, I just don't want to see it on my view - dont want 'sources' directly. i want my notes, which will be interleaved with links i drop in. the AI should index the content of those links as needed, as well as my ideas, as the default. (i know i can make notes sources but its clunkier imo) - would be amazing if notebookLM could parse gemini chats when i drop those links into my notes. a lot of the time i'll talk to gemini to research something, then throw the chat link into obsidian so I can reference it later, but there's no existing solution for me to be able to later talk to an AI that has background context of a set of chats - finally, want to be able to share my entire set of notes on a premade UI like Quartz, along with the git history sorted by most recent first, so that people can easily understand what I'm working on. even better if they have access to the AI chat as well to interact with my process. There's also a ton of more 'vision' features I want, orbiting around CIRL/interactive learning, passive context collection, making the answering AI more agentic, social networking around the shared context/ideas, agent to agent collaboration, but that comes later. As a simple example, giving the AI a secure VM that it could code up small experiments on. I'll often a read a paper and want to implement a simple version of the ideas but the friction is so high. And obviously I'd want this to be auto-shared. Overall, I'm trying to maximally expose the process of my work so that AI can be maximally useful to me, now and in the future. So much friction associated with getting my granular context persisted and legible in real time. I think Obsidian does decent with note taking, sharing, and git persistence if you set it up, while NotebookLM does well with AI augmentation and link parsing. At a high level, I think what's needed is largely an Obsidian frontend but NotebookLM backend.

i need something halfway between @obsdmd and @NotebookLM who is building this

in the early stages, market froth is indistinguishable from RSI

Waiting to see if this is actually rolled out or not, but 12M context would be a timeline decrease for sure

Alexander Whedon @alex_whedon

2 weeks ago

Here is the technical report on SubQ 1.1 Small. subq.ai/subq-1-1-small… This is the second iteration on our Subquadratic Sparse Attention (SSA) model, and the first to be deployed with design partners in the coming weeks. The results are compelling and verified by

66 78 551 142K 276

RSI but for whatever you care about

will brown @willccbb

2 weeks ago

been beating this drum since early 2025, seems like people are starting to see why it's so important :) RL works -> "train or get trained on" -> open models + post-training infra are the path to institutional flywheels + democratization of AI progress

keyboard_arrow_left Previous keyboard_arrow_right Next

6 28 392 36K 251

Can’t get over the fact that enterprises “doing their own RL” feels like the equivalent to businesses “building their own railroads”. One difference is obviously that open source models have no railroad equivalent. Another difference is that my own railroad would at best likely be the same as an existing railroad, whereas custom RL promises to improve performance.

elie @eliebakouch

4 weeks ago

subagents, teams of agents etc. will be first class citizens soon (if not already) two things here: 1) you want to maximize token efficiency even more 2) training/serving on your own harness gives you an even bigger boost than before benchmarks in the opus 4.8 model card show

4 8 79 7K 38

@sethkarten thank you for your wisdom

@tenobrus what is this from?