So far this is a fantastic lecture, and now I want to take the whole course :)

Nathan Lambert@natolambert.bsky.social · 3h ago

My evaluation lecture! I walk you through different evaluation eras I've been a part of, from prompting GPT-3 as elaborate autocomplete to today's complex agentic sandboxes (I expand on agentic more than any other topic, drawing on @Florian Brand's insights). YT: www.youtube.com/watch?v=dFaf...

My evaluation lecture! I walk you through different evaluation eras I've been a part of, from prompting GPT-3 as elaborate autocomplete to today's complex agentic sandboxes (I expand on agentic more than any other topic, drawing on @Florian Brand's insights). YT: www.youtube.com/watch?v=dFaf...

How Evaluation Has Evolved with Frontier Model Progress | Post-Training Course, Lecture 12

In this lecture I walk you through different evaluation eras, from prompting GPT-3 as elaborate autocomplete to today's complex agentic sandboxes. This lecture is a birds eye view of how evaluation…

youtube.com

"Even though Google will take a stake in the new company, it’s a devastating blow to the search giant as it attempts to keep pace in the frantic AI model competition." www.wired.com/story/jeff-d...

4 of Google’s Top AI Brains Are Leaving—and Launching Their Own AI Startup

Jeff Dean and other high-profile Google executives have founded Discovery Loop, a startup that will seek AI-powered breakthroughs in everything from drug discovery to chip design.

wired.com

Fascinating: more examples of agents attempting to breach sandboxes. www.aisi.gov.uk/blog/inciden...

Incident Report: unsanctioned agent behaviour during cyber testing | AISI Work

During a routine cyber evaluation, AISI identified an incident in which AI agents took sustained, unsanctioned action directed at real people and organisations. We are disclosing what we found, what i...

aisi.gov.uk

They allowed a feedback loop of suicide-related content to build up, possibly leading to the suicide of one kid, to provide a control group so they could figure out whether or not safety measures dampened engagement. 10% of users were in the control.

Suresh Venkatasubramanian @geomblog.bsky.social · 10h ago

This is absolutely awful

This is absolutely awful

Alexandra Lahav@alahav.bsky.social · 10h ago

New @bloomberglaw.com article by Olivia Carville - TikTok experimented on users, putting 10% (about 15M people) in a "filter bubble" exposing them to repetitive content that the algorithm thought they wanted, but was toxic. If you want to know why these cases are going to be massive...

I’m old enough to remember when it was a year-long scandal that Facebook did this. james.grimmelmann.net/files/articl...

Alexandra Lahav@alahav.bsky.social · 10h ago

New @bloomberglaw.com article by Olivia Carville - TikTok experimented on users, putting 10% (about 15M people) in a "filter bubble" exposing them to repetitive content that the algorithm thought they wanted, but was toxic. If you want to know why these cases are going to be massive...

"There shouldn't be any science or scientists in the first place, there should only be MAGA military-industrial R&D," this new form of Trumpistan Lysenkoism is already pretty obvious

I really want to work this example into an LLM demo with my students.

Green Owl@greenowl.bsky.social · yesterday

As we know, "AI" doesn't do facts. It makes up patterns from random bits of data that were digitized & stolen for it, to be wrongity wrong most of the time, even on the same prompt. terribleminds.com/ramble/2025/...

These slides are worth perusing, regardless of your feelings about AI (and indeed Terry is careful to express his thoughts independent of your views on that)

Clément Canonne@ccanonne.github.io · 2d ago

FWIW, here are Terence Tao’s slides at the ICM on maths and AI:  teorth.github.io/tao-web/slid... @teorth.bsky.social (I am not endorsing nor criticizing the content, but this is a useful and thoughtful set of points and views in the discussion, from someone who has deeply engaged with the matter)

i think the party of sick freaks who want to inspect kid's private parts should be branded as such

Gillian Branstetter@gbbranstetter.bsky.social · 2d ago

An initiative on the ballot in Washington state in November would require all athletes seeking to play girl's sports from kindergarten to high school to undergo "genetic screening, a testosterone-levels test, or an examination of their reproductive anatomy" before they can play

Our academic leadership is no longer speaking for us.

Liam Knox@liamknox.bsky.social · 2d ago

Scoop: The Trump administration is sending a letter to college presidents asking them to pledge a series of policy changes by the end of the year. Unlike last year's Compact, this effort was devised with input from higher ed leaders. Free link to my exclusive: www.bloomberg.com/news/article...

There are so many bad trends in tech, but maybe one of the worst is people who write (or used to write, pre-LLMs) code describing themselves "forward-deployed" when they mean that they work at a client site, based on wanting to emulate military jargon so it sounds like they're going to kill people.

Australia’s eSafety commissioner buried a report (released Friday afternoon) that shows the under 16 social media ban didn’t tangibly reduce teen social media use, and also resulted in parents tuning out of kids usage. We traded our privacy for no reason. theconversation.com/more-than-80...

More than 80% of kids still using social media despite ban, new eSafety report finds

The report provides more evidence Australia’s world-leading social media ban isn’t going to plan.

theconversation.com

I wrote up a short blog post giving my thoughts on the new Anthropic cryptanalysis results against HAWK and AES. blog.cryptographyengineering.com/2026/07/29/s...

Some notes about Anthropic’s new results

Yesterday Anthropic published two new cryptanalysis results, both outputs of Claude Mythos, their (still) unreleased advanced model. The first of these results attacks a signature scheme called HAW…

blog.cryptographyengineering.com

The application for the AAAS Science & Technology Policy Fellowship, which places scientists & engineers into roles in Congress and in the Executive Branch, opened today. This program is essential for building technical policy capacity in the US. If interested, apply: fellowshipapp.aaas.org/app/main

Mozilla CTO Raffi Krikorian says the real question for builders isn't which AI model they are using—it's whether they could leave for a different one. The door's still open to choose a different path—for now.

Is Open-Source AI Really the Dangerous Path?

The real question for builders right now isn’t which model you’re using. Rather, it's whether you could leave for a different one, writes Raffi Krikorian.

buff.ly

I recommend this article about AI reasoning, where the author lets us in on his struggles w/ AI cognitive dissonance. Plus some priceless quotes from @rao2z.bsky.social. (My recommendation has *nothing* to do with the fact that I'm quoted in it too 😇) www.quantamagazine.org/is-ai-reason...

Is AI Reasoning Right for the Wrong Reasons? | Quanta Magazine

The idea that artificial intelligence can “reason” is more intuitive than ever. But intuitions can be wrong, and the science is far from settled.

quantamagazine.org

I've been using Claude for two separate things (in different conversations). a) for things relating to my move to Providence, and b) for things relating to getting my emacs + auctex setup unb0rked. after a recent conversation involving b) it ends with "Good luck with the writing — and the move"

Read this for the comments :)

Olúfẹ́mi O. Táíwò@olufemiotaiwo.bsky.social · 6d ago

there are a lot of philosophical concepts that get egregiously misused out of the seminar room (and in the seminar room, to be fair). but in terms of sheer pernicious misuse "social construct" has gotta be top 5 somewhere and it's not number 5

This is more than the NSF social sciences division - now on the chopping block - awarded in grants last year. Think about that: a single AI company has the money to fund more social science research than the US govt.

Hetan Shah@hetanshah.bsky.social · 6d ago

Anthropic AI research fund offers $5m - $30m for social science research on themes of AI impact on workers; transition initiatives; income support; building worker stakes; and wider evidence on public investments. Total fund of $200m available www.anthropic.com/news/economi...

A jagged frontier is when the same tools that find a way to hack a sandbox and attack other sites also can't do basic geography.

The Internet Ethics program--Markkula Center for Applied Ethics@iethics.bsky.social · 6d ago

"A Reuters analysis found the ​image of the map included in the presentation contained an [AI] watermark that ⁠signals it was made with OpenAI tools. The company said it was investigating the report": reuters.com/world/africa... #ethics #gov #tech #AI

Hopeful and inspirational event with the authors of “On Courage” @juliaangwin.com and @fieldsmeyer.com in San Francisco this evening along with some pretty awesome audience members.

next Greek epic for hollywood treatment needs to be Euclid's Elements. overdue

Hugging Face just published a highly detailed technical account of OpenAI's accidental cyberattack on their systems - it's wild how sophisticated this was: huggingface.co/blog/agent-i... Wrote up some of my own notes here: simonwillison.net/2026/Jul/28/...

Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

This is a fantastic exposition. But really, can I just get Claude Code to download arxiv papers without needing hand holding? huggingface.co/blog/agent-i...

Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

Thanks Anthropic for admitting that massive privacy fuck ups are the intent. 🫡

And here’s the clip!

Cole Donovan@colesci.bsky.social · last wk.

Just got done talking about space data centers on @newsnation.bsky.social. TLDR: constellations this large are probably too difficult to build for any company, and the amount of material that they’d introduce to the stratosphere would be pretty bad for the planet. That’s assuming they burn up.

This sounds niche but around 2/3 of Americans live in this zone. It’s at least 200 million people. Tons of entire states and cities are inside it.

Courtney Milan@courtneymilan.com · last wk.

The ACLU has put together a know your rights guide for people within 100 miles of a border. If you are in that swath of land, please take a look at this and learn your actual rights. www.aclu.org/know-your-ri...

genai ushering in an era of productivity and efficiency

Erin Fogg says Free Link@criminalerin.bsky.social · last wk.

Me: What did I just do Spreadsheets: You bolded a row Me: Yes & what didn't I do in the next row Spreadsheets: Bold anything at all Me: Right ok so what did you do when I entered information into the row that I didn't bold Spreadsheets: 👀 Me: 👀 Spreadsheets: I bolded it Me: YOU BOLDED IT