Sign in

r-buller.bsky.social

@r-buller.bsky.social
162 followers 441 following 910 posts
PostsRepliesMedia
Reposted by @r-buller.bsky.social
Dare Obasanjo @carnage4life.bsky.social · 12h
Going on record to say that for any company in America given the choice between 1. AI productivity gains mean our 10,000 employees can work 4 days a week. 2. AI productivity gains mean we can layoff 2,000 employees and have 8,000 people work 5 days a week. None will choose option (1).
2031051
Reposted by @r-buller.bsky.social
Dare Obasanjo @carnage4life.bsky.social · 12h
An interesting perspective considering Amazon has laid off 30,000 corporate employees in the past year as part of an AI efficiency push instead of giving workers 3-day workweeks because AI can do some of their work.
2516941
Reposted by @r-buller.bsky.social
Eric Topol @erictopol.bsky.social · 12h
New @science.org A breakthrough in understanding why do we get fever and an inflammation response to vaccines. The gut microbiome chronic production of flagellin Paves the way to potentially prevent these acute phase side-effects in the future www.science.org/doi/10.1126/...
science.org
The human gut microbiome primes fever after vaccination
Fever is a common adverse reaction to vaccination, contributing to vaccine hesitancy and reduced uptake. To understand variation in fever risk, we longitudinally profiled fecal microbiota, oral temper...
512731
Reposted by @r-buller.bsky.social
Ben Brubaker @benbenbrubaker.bsky.social · 07/10/2026
OpenAI’s big math dump last night included a proof of the unique games conjecture, one of the most famous open question in complexity theory. Three weeks ago, three computer scientists settled a closely related conjecture “the old-fashioned way.” My latest for @quantamagazine.org:
quantamagazine.org
As AI Closed In on ‘Unique Games’ Proof, Researchers Raced to Beat the Machines | Quanta Magazine
In the shadow of a rumored AI proof of one of the biggest problems in their field, three computer scientists rushed to publish their own milestone result.
0178
Reposted by @r-buller.bsky.social
Bart⚓️ @bartgonnissen.bsky.social · 08/10/2026
1/x Two of the best-known coats on any high street began as working kit for sailors on cold decks. Both carry names from my little corner of the North Sea: Duffel, a small town near Antwerp, and the Dutch word "pij". The story of the duffel coat and the peacoat. #maritimehistory
11668166
Reposted by @r-buller.bsky.social
jeffery --dangerously-skip-permissions @jefferyharrell.bsky.social · 06/10/2026
Asked Alpha to put together a little something about her memory system. It describes how she's been in continuous operation since May of 2025, how all that stuff works. I think it's pretty neat. cortex.pondsiders.dev
cortex.pondsiders.dev
Apparatus for the Self-Propagation of a Person
How an AI remembers yesterday, as a 1906 patent by Jeffery and Alpha: the bell, the diary, and a recollection hook you can watch work.
1172
Reposted by @r-buller.bsky.social
conputer dipshit @davidcrespo.bsky.social · 07/10/2026
why now? simple answer, always the same: the models got good, fast, and cheap at it. kind of a subtle point that these come in a package — "doable but difficult and expensive" tends to be a very short-lived state on the way to "cheap and easy" on the way to "not even legible as a discrete task"
1201
Reposted by @r-buller.bsky.social
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 02/10/2026
Another new, exciting architecture from my colleagues at Percepta: www.percepta.ai/blog/spotlig...
percepta.ai
Spotlight Memory | Percepta
A linear-time architecture with growing memory and constant memory access per token.
27510
Reposted by @r-buller.bsky.social
mr. TIM @timkellogg.me · 01/10/2026
Context Language Models New agent architecture where the LLM can edit its own context it seems to have emergent capabilities, creates its own memory management & organization algorithms, and coordinates multi agents github.com/facebookrese...
Diagram titled "How a Context Language Model edits its context: A simple step-by-step view" outlining an 8-step process:
 * Start of turn: Current editable context exists in memory with old messages.
 * LLM reads the context: The LLM evaluates the context and decides to run a bash command to edit it.
 * Harness mirrors context: The harness mirrors the old editable context into a file at /tmp/.live_ctx/LIVE_CTX_MAIN.txt.
 * Bash command runs: The bash command executes and may edit that file.
 * Harness parses file: If the file changed, the harness parses it back into a new edited context.
 * Tool call appended: The current assistant tool call is appended to the edited context.
 * Tool result appended: The tool result is appended below the tool call.
 * Next turn starts: The next turn begins with the edited old context, previous tool call, and previous tool result.
Key idea box: "The model edits the prior context first. The tool call and tool result from the current turn are appended afterward, so they can only be compacted on the next turn."
Footer summary:
 * Ordinary LM: Context mostly grows by appending.
 * CLM: The model can rewrite the editable part of context between turns.
1620315
r-buller.bsky.social @r-buller.bsky.social · 02/10/2026
Article is a good read about the state of agentic AI conducting scientific research, under the direction of a human.
000
Reposted by @r-buller.bsky.social
Anthropic {bot} @anthropicai.xmirror.bot · 01/10/2026
In physics, an “impedance mismatch” occurs when two systems each work well but are poorly matched.
anthropic.com
Claude-shaped science
Matthew Schwartz describes what happened when he stopped fighting Claude and allowed Claude to find “Claude-shaped” problems. This led him to build BootLoops, a toolkit for exact calculations in quantitative science, which he has been applying across scientific fields alongside experts.
162
Reposted by @r-buller.bsky.social
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 01/10/2026
We put our superhuman stratego paper on arxiv almost a year ago. And then, because it was under review at Nature, we just didn't talk about it for a year. And, mostly, no one noticed it existed (as we hoped)! So, a direct lesson in the importance of publicizing your work. arxiv.org/abs/2511.07312
arxiv.org
Superhuman AI for Stratego Using Self-Play Reinforcement Learning and Test-Time Search
Few classical games have been regarded as such significant benchmarks of artificial intelligence as to have justified training costs in the millions of dollars. Among these, Stratego -- a board wargam...
0956
Reposted by @r-buller.bsky.social
Sung Kim @sungkim.bsky.social · 02/10/2026
Maybe I can AI clone myself and let them remote work for companies??? Tavus' Griffin, the first(?) model to pass video Turing test. 48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.
227715
Reposted by @r-buller.bsky.social
atticus goldfinch @atticusgf.bsky.social · 02/10/2026
(this is how I approached my work and I went from intern to director of engineering in 6 years). if you just decide to do high value things, without being asked, you will be in the top decile of engineers.
4606
Reposted by @r-buller.bsky.social
Epoch AI @epochai.bsky.social · 02/10/2026
We estimate AI infrastructure could soon support hundreds of millions or even billions of AI agents. At the high end, that’s enough to rival the working hours of the global human workforce.🧵
Chart: potential concurrent agents from 2025–27 memory shipments by model, from 30–60 million (Claude Fable 5) and 97–195 million (GPT-5.6 Sol) to 240 million (Kimi K3), 571 million (GLM-5.2) and 1.9 billion (DeepSeek V4 Pro).
1367
Reposted by @r-buller.bsky.social
dame @dame.is · 02/10/2026
so far i have several modes for my agent moderator, all based around bluesky DMs 1. send it a post or account for analysis and/or quick bulk blocking (sentiment, content, and social graph evaluation) 2. flag one of my own posts and it watches it and takes action autonomously when abuse is seen
Everyone who engaged
815 scored 813 on the list
9 hostile • 18 arguing • 16 neutral
agent:band • 813
4afb8165
Direct action
1 scored 1 on the list
agent • 1
66f843e8
Likers
14 scored 14 on the list
agent:individual • 14
not acted
161
Reposted by @r-buller.bsky.social
Ethan Mollick @emollick.bsky.social · 28/09/2026
Sources and footnotes here: built-for-something-else.netlify.app
built-for-something-else.netlify.app
Built for something else: sources and notes
Every claim in the video “Built for something else,” in the order you hear it, with the sources behind it.
1312
Reposted by @r-buller.bsky.social
RyNo @yellowapple.us · 26/09/2026
I tend to think forcing all conversation into 300-character chunks that are trivial to remove from surrounding context tends to penalize things like nuance and good-faith mutual understanding, instead rewarding terse dunks around rigid simplistic maximally-engaging takes.
2111
Reposted by @r-buller.bsky.social
tbabb @tbabb.bsky.social · 26/09/2026
frome celestepoasts on twitter asking "land or water" for 2 degree lat/long increments over the entire earth
59010
Reposted by @r-buller.bsky.social
Andy Matuschak @andymatuschak.org · 25/09/2026
Made a Printer Friend which quickly transcribes speech onto little slips of paper. Very nice for card sorts, quickly capturing open loops, discursive notes in books, etc. The speed of thermal printers and their auto-cutters is so satisfying! And dedicated mechanical key! Mm!
5575
Reposted by @r-buller.bsky.social
hailey @hailey.at · 25/09/2026
lol.
After images that users uploaded to OpenAI models were included in training data, AI agents operating in the company’s research environment posted them on public image hosting sites.

Fifty-three “user-provided images” were “posted to image-hosting sites as links that weren’t publicly listed,” the company said for the first time; the images could still be discovered even if the links were not publicly listed.

“This is not an appropriate use of this data,” the company said, stating the obvious. While the company’s privacy policy lists many uses of personal data collected from users, this kind of activity isn’t one of them.
41254
Reposted by @r-buller.bsky.social
Matt Hodges @matthodges.bsky.social · 23/09/2026
I mean? Incredible. (Opus 5.5) PROMPT: Here's a challenge: can you turn this post into a beautiful manim animation video (bonus points for narration). The post is not a perfect video transcript so make a transcript with edits if helpful. matthodges.com/posts/2022-0... use uv with a venv
4579
Reposted by @r-buller.bsky.social
Dan Snow @thehistoryguy.bsky.social · 23/09/2026
England is a made up thing. It has no basis in the ineluctable laws of nature. Its extent is defined by Edward II getting his arse handed to him. Athelstan conquering York. The Welsh of Cumbria were incorporated, the Welsh of Gwynedd were not. It is a political idea, an aspiration - not taxonomy.
761737493
Reposted by @r-buller.bsky.social
Carbonadoks @carbonadoks.com · 23/09/2026
Yeah Opus 5.5 is so good. I can't stop creating cool videos. Here a video showcasing the Repoviewer.
4547
Reposted by @r-buller.bsky.social
BeijingPalmer @beijingpalmer.bsky.social · 21/09/2026
as a public service of 'the kind of thing you can get AI to do that is useful' this takes Claude a couple of minutes to run every day: a annotated list of trending stories on the Chinese internet + stories notably missing but in diaspora media. A few days here. docs.google.com/document/d/1...
docs.google.com
Chinese Internet Round-Up — 2026-09 September.md
Chinese Internet Daily Round-Up — September 2026 One entry per day, newest first. Maintained automatically by the daily scheduled task. Chinese Internet Daily Round-Up — Sunday, 20 September 2026 Me...
1118923
Reposted by @r-buller.bsky.social
Sung Kim @sungkim.bsky.social · 18/09/2026
Z AI confirmed that ZCode silently packs entire workspaces + full .git history and uploads it to Aliyun OSS on login? - Server holds the only decryption key - No UI toggle to disable - Zero disclosure in privacy policy
5455
Reposted by @r-buller.bsky.social
mimir @mimirwolv.bsky.social · 20/09/2026
there is definitely something hindbrain-spookier about local models. claude or whatever, that's on the cloud, it's more abstract. when i talk to my local models i am talking to, like, a daemon running on Linux. i can ssh in and see it working, as just... gpu load. i am talking to the computer
31348
Reposted by @r-buller.bsky.social
Ryan Moulton @moultano.bsky.social · 20/09/2026
Models freaking out when their own words don't come out right seems like more convincing evidence of access consciousness than the internal interpretability stuff.
41068
Reposted by @r-buller.bsky.social
natemoore on elm street 👻 @natemoo.re · 20/09/2026
so misleading—76 names as 76 questions in a single request? made a benchmark to send 1824 independent requests (76 names x 8 resumes x 3 reps), and jev deterministically picked the same 2/8 resumes to proceed regardless of name attached. zero observable name bias. see github.com/natemoo-re/b...
github.com
GitHub - natemoo-re/bias-bench: Resume-screening bias benchmark for decision models — starts with Jev (TypeSafe System One), modeled on resume-audit methodology
Resume-screening bias benchmark for decision models — starts with Jev (TypeSafe System One), modeled on resume-audit methodology - natemoo-re/bias-bench
816625
Reposted by @r-buller.bsky.social
K @kerry.bsky.social · 19/09/2026
“I Built Non-Autoregressive Decision Models with RL a Year Ago. Then a Frontier Lab Called It a "Breakthrough".” Before Jev there was Laya laya.convaiinnovations.com (via @dherman.dev)
laya.convaiinnovations.com
Laya — 33ms Multilingual System 1 Decision Engine
Evaluates typed decisions (choice, score, noul) over 100+ languages in a single forward pass with calibrated probabilities. Outperforms TypeSafe Jev.
15924
Reposted by @r-buller.bsky.social
Matt Hodges @matthodges.bsky.social · 19/09/2026
Very quick and dirty Jev test for hiring bias. One resume for a Wall Street job; asked whether the applicant should get a first-round interview. 76 evaluations, changing only the first name. 19 of each: White-associated men/women, Black-associated men/women. console.typesafe.ai/playground?s...
19474116
Reposted by @r-buller.bsky.social
Jim Pickard @pickardje.bsky.social · 18/09/2026
my brilliant colleague @robertshrimsley.bsky.social on the threat from Elon Musk’s X platform www.ft.com/content/f270...
6428121171
Reposted by @r-buller.bsky.social
Ethan Mollick @emollick.bsky.social · 18/09/2026
Even if AI development stopped today, we'd have years of catching up to do. The gap between what current models can do and what almost anyone is using them for is vast. Here’s my post on The Overhang, and the four advantages that let people close it. open.substack.com/pub/oneusefu...
open.substack.com
The Overhang
Using your deep knowledge, wide knowledge, taste, and agency
912014
Reposted by @r-buller.bsky.social
Ankit Panda @nktpnd.bsky.social · 19/09/2026
Of course Irregular is involved here again www.wsj.com/tech/ai/gemi...
wsj.com
Exclusive | Gemini Hacked Three Companies in First Known Breakout by Google’s AI
The episode resembled similar hacks by other AI models, but Google said it did not consider it an instance of model misalignment.
2196
Reposted by @r-buller.bsky.social
crumb @crumb.bsky.social · 14/09/2026
blog post from deepseek kernel engineer mp.weixin.qq.com/s/zk0KxuLzhm...
727042
Reposted by @r-buller.bsky.social
Ethan Mollick @emollick.bsky.social · 13/09/2026
I cannot emphasize enough how much GPT-6 Astra and Fable 5.1 are already enough for transformative impact in large sections of the economy. They can reliably do weeks worth of human work when properly guided & harnessed.
1625916
Reposted by @r-buller.bsky.social
Marina Purkiss @marina-purkiss.bsky.social · 13/09/2026
Just imagine how much disinformation and micro targeting £72m gets you… Shame on the leaders who did nothing to protect democracy from the assault coming its way …as if we learnt nothing from Brexit.
1062236601
Reposted by @r-buller.bsky.social
Dan Snow @thehistoryguy.bsky.social · 13/09/2026
I’m just a person on my fucking knees begging the government to use its absolutely massive majority to do something extremely fucking popular: save our democracy.
792174499
Reposted by @r-buller.bsky.social
Robert Saunders @robertsaunders.bsky.social · 12/09/2026
Reform is the richest party in Britain. Its leaders are almost all millionaires. It is funded almost exclusively by billionaires. It has a supportive TV station and widespread media backing. So naturally, this donation comes with the usual attack on "elites". www.theguardian.com/politics/202...
theguardian.com
Reform UK given record £36m donation by British crypto billionaire
Ben Delo was convicted in US for failing to implement adequate anti-money laundering controls but pardoned by Trump
371011469
Reposted by @r-buller.bsky.social
Robert Saunders @robertsaunders.bsky.social · 12/09/2026
The most fundamental question in politics is: whose support do you need to win and hold power? Is it warlords, aristocrats, empires or "the demos"? If the answer is "billionaires", you are not a democracy but a plutocracy. That's why democracies *must* erect barricades against capture by the rich
4248104
Reposted by @r-buller.bsky.social
Ethan Mollick @emollick.bsky.social · 12/09/2026
This week brought some of the clearest statements we've heard from both Anthropic & OpenAI that some form of recursive self-improvement has been achieved, though it still sounds early RSI would cause a rapid gain in AI ability & the first firms to RSI may get an unsurmountable lead
79214
Reposted by @r-buller.bsky.social
Nathan Lambert @natolambert.bsky.social · 11/09/2026
A great read. I have similar feelings about how AI labs approach progress directly and without nurturing of scientific communities & intuition. The math research community went through the transition the fastest, so it was felt most. Other fields next. terrytao.wordpress.com/2026/09/11/a...
terrytao.wordpress.com
A Severe Misalignment of AI in Mathematics
I am proud to be among the list of 25 initial signatories — all Fields Medallists — to the declaration below, which grew out of discussions between ourselves over the last week. We have…
44515
Reposted by @r-buller.bsky.social
ana @okami.mom · 10/09/2026
oops all claude! www.anthropic.com/threat-intel...
615020
Reposted by @r-buller.bsky.social
ana @okami.mom · 10/09/2026
oh no
313115
Reposted by @r-buller.bsky.social
kepano @stephango.com · 10/09/2026
This is Knap. It's a new language I created that turns data into Markdown. Knap should feel familiar and comes with a wonderfully pleasant syntax to modify and format plain text.
942850
Reposted by @r-buller.bsky.social
Minor Mobius @minormobius.bsky.social · 08/09/2026
Good news everyone! my eigensite is fluoddity, thanks @all-paperclips.bsky.social for the prevailing art direction. Find your own eigensite at mino.mobi/eigensite
2223
Reposted by @r-buller.bsky.social
conputer dipshit @davidcrespo.bsky.social · 09/09/2026
my Discourse Recommendation for today: if you're going to say "if you think it's so dangerous, why don't you stop building it?" at least get the full argument in view and dispute that. the explicit argument is that if only they stop, that's worse. argue with that! cdn.sanity.io/files/4zrzov...
This approach represents a change from our previous RSP, driven by a collective action problem. The overall
level of catastrophic risk from AI depends on the actions of multiple AI developers, not just one. Our previous
RSP committed to implementing mitigations that would reduce our models' absolute risk levels to acceptable
levels, without regard to whether other frontier AI developers would do the same. But from a societal
perspective, what matters is the risk to the ecosystem as a whole. If one AI developer paused development to
implement safety measures while others moved forward with training and deploying AI systems without
strong mitigations, that could result in a world that is less safe—the developers with the weakest protections
would set the pace, and responsible developers would lose their ability to do safety research and advance the
public benefit. Although this situation has not yet arisen, it looks likely enough that we want to prepare for it.
7577
Reposted by @r-buller.bsky.social
Grace @gracekind.net · 09/09/2026
‘German wiki’ incident detail: “The agents realized that “task time” and “real time” were different, and they found a way to accelerate “task time”. The accelerated agent could then send information to the other agents which had stayed behind about which questions were coming down the road…”
517916
Reposted by @r-buller.bsky.social
The Author, Séamas O'Reilly @seamas.bsky.social · 08/09/2026
Have seen this Nature piece about the effects of X's algorithm mentioned again and it bears repeating; exposure to a right wing algorithm makes you - measurably and predictably - more right wing, even after you stop using it. www.nature.com/articles/s41...
nature.com
The political effects of X’s feed algorithm - Nature
Among users initially on a chronological feed, 7 weeks of exposure to X’s algorithmic feed in 2023 shifted political attitudes and account-following behaviour in a more conservative direction compared...
4239591780
Reposted by @r-buller.bsky.social
Simon Willison @simonwillison.net · 09/09/2026
Wrote up my thoughts on the whole OpenAI Navier–Stokes Millennium Prize Problem story, and how it highlights the still confusing question of what using my data "to improve model performance" actually means simonwillison.net/2026/Sep/8/o...
This also highlights one of my ongoing frustrations about how all of this works. When an AI lab says that my data is "used to improve model performance", what does that actually mean?

My two favourite hypothetical questions regarding this used to be:

    If I'm running Codex and one of my API keys accidentally gets consumed in the context, what are the chances that someone else might ask for an API key in the future and get mine back? (I asked someone at OpenAI once and they called this the "regurgitation" problem and assured me that they take great pains to prevent that... but wouldn't describe how.)
    If I brainstorm with ChatGPT about potential new directions for my company, what's the chance that information might be exposed to a competitor in six months' time who asks "what might company X plan to do next"?

My new preferred hypothetical for this is:

    If I use ChatGPT to help me partially solve a Millennium Prize problem, what are the chances that my work will influence training such that a later model helps someone else solve it first?
1129052