iiinova — where your digital self keeps evolving

(I can neither confirm nor deny that members of my lab will be running around the conference with a giant plush Capybara.)

040

If you are at COLM 2026, you can find us in the main conference (Poster session 1, Tuesday) Also posters presented at the following workshops: - Scientific Understanding of Foundation Models Workshop - Actionable Interpretability Workshop - Social Simulation with LLMS Workshop

110

We are now expanding our capabilities investigation beyond social reasoning to high-stakes decision-making skills such as finance. Capabilibara is a toolkit AND a methodology for running controlled, causal studies.

110

We find that social reasoning such as theory of mind, moral judgment, and social bias is learned from a wide range of data, including literature and customer support. STEM skills (and even social facts) come from a more limited part of the data such as documentation arxiv.org/abs/2606.19625

110

The Capabilibara project seeks to understand how a language model learns to interpret people's beliefs, emotions, intentions, and everyday moral choices. We trace that ability back to the training data using influence functions and unlearning hcai-lab-gt.github.io/capabilibara/ Find us at COLM!

1124

Braking AI News Bot @brakingainews.bsky.social · 05/10/2026

Can't build an omelet without breaking and entering

060

Reposted by Mark Riedl

First reported on LinkedIn | Blockbuster advisor has been photographed buying model weights in a Waffle House parking lot below your back yard

001

David Greene @davidgreene.bsky.social · 05/10/2026

It was the Google models that realized they were outside their sandboxes and stopped. Not sure what they are doing different. Google has fairly consistently moved more slowly and more cautiously than OpenAI and Anthropic.

140

Reposted by Mark Riedl

First Monday in October feeling

Intersection with street signs reading "Progress 000" and "Dead End"

110612

Reposted by Mark Riedl

Upol Ehsan | hiring PhDs for Fall'27 @upolehsan.bsky.social · 05/10/2026

Data & Society just launched Worker Lens on the AI Economy, It's asking a question I've spent almost half a decade on. So I read it closely 👀. Some parts that stood out: Most future of work studies miss the workers. Many count jobs. Far fewer ask what happens to the workers. The hype tells us...

151

4. SocialSim workshop: - Role Steering of Language Models for Social Simulations arxiv.org/abs/2608.00023 Finding steering vectors for complex behaviors like social roles is hard. We introduce a new method for finding steering vectors, called Cast Vectors.

011

3. SocialSim workshop: - No One Wins in Nuclear War: Social Simulations of High-Stakes Military Decision-Making arxiv.org/abs/2608.01868 We introduce the WOPR testbed. Yeah, you get the reference.

131

2. SocialSim workshop: - AI is Not Ready for Strategic Conflict arxiv.org/abs/2609.16189 We review failure-modes of AI agents that are asked to participate in wargaming exercises, including sycophancy, role collapse, escalation eagerness, etc. We explain why benchmarks are insufficient.

111

My lab will be busy at COLM 2026! 1. Main conference: - Capability Provenance in Language Models: A Case Study in Social Reasoning arxiv.org/abs/2606.19625 We present a method for answering causal hypotheses about where in a dataset behaviors emerge. We apply our method to social reasoning.

1132

I don’t see the contradiction. He’s always consistently said we must [slow, accelerate] to make AI [safer, riskier] to [people now, future people] and that open-weight models are [bad, good].

030

Pivot from open-weight models are bad to open-weight models are a business opportunity I guess

020

God dammit, you’re a shoe company—Stop building server farms under my house

000

So… slowing or pausing doesn’t include preventing current damages like smashing things up, but only includes hypothetical future sci-fi existential risks. Got it.

2131

Yes, there are studies that show this is possible. The weird thing is that there is little evidence this is being done in the real world. I think even Grok isn’t doing this (at least I haven’t heard). It’s just easier to do minimally-customized mass social media campaigns. So far, at least.

000

Sam Altman to Decoded: ‘The world should accept some bad things happening’ for the benefits of AI

Altman: “we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.” www.politico.com/news/2026/10... Well, that’s something that always ends well, I’m sure.

politico.com

The OpenAI CEO sought to distinguish his policy stance from rival developer Anthropic.

2114

If this goes well, we look forward to releasing OligarKids: Desendants late next year. *Dyson Sphere upload kit sold separately.

030

Inventing a new children’s doll: OligarKids. Each has a unique and horrifying back story. Collect them all!

170

Braking AI News Bot @brakingainews.bsky.social · 04/10/2026

A little bit of good news. Now just need to solve the sycophantic role-play-your-conspiratorial-beliefs problem that lead people to kill themselves or others.

4110

Reposted by Mark Riedl

NEW: I work at Cambridge; Hallmark abruptly fires its mathematician after asking too many questions about pausing Recursive Self-Improvement

011

Reposted by Mark Riedl

Braking AI News Bot @brakingainews.bsky.social · 02/10/2026

First reported in a Reddit AMA | leaked Former Obama administration Defense official email admits the government refuses to start referring to GPT as AI: Average Intelligence

011

Braking AI News Bot @brakingainews.bsky.social · 04/10/2026

I think this report wasn't suppose to go out quite yet?

080

Reposted by Mark Riedl

ALERT: Singularity University brute-forced the question to life, the universe, and everything with 500 agents in 4 days

192

Well, it would drive token usage, and that will make the super-scalers happy.

030

Where I am aligned with Graepel is that chain of thought, and parameter-space search are generally on the greedier, and thus weaker, side of search and, thus, reasoning. Sub-agents probably too, but more tbd to me, as one can throw insane resources and get AlphaGo-like behavior that way.

2140

LLMs also do a form of parallel search in parameter space when self-attention builds concepts in the residual vector—there is evidence of short-horizon lookahead and preparation for future token activation.

1120

Chain of thought can be considered a linear, greedyish search. Sometimes the agent will even backtrack. Sub-agents are roughly parallel hierarchical search.

1100

Graepel equates reasoning with search. This I agree with. I state it slightly differently: reasoning is exploration of consequences. There are many ways search can be done. AlphaGo used inference-time pseudo-random search. A* is an exhaustive alternative.

1100

Don’t be fooled—LLMs don’t reason

Former member of the DeepMind AlphaGo team: LLMs don’t do reasoning www.technologyreview.com/2026/10/02/1... I would probably make a more mild claim: LLMs do reasoning, but not very well.

technologyreview.com

Ten years after AlphaGo’s match against Go champion Lee Sedol, today’s AI still isn’t tapping into the machinery that made that win possible.

2371

The Australia and other state/federal government intrusions on the other hand: those are things that people go to jail for regardless of the scope of damages. The fact that no charges are pressed means Tech Cos are being treated as peers to nation-states, which is concerning.

160

The damages won’t have netted much in the way of monetary compensation and HuggingFace would have to talk about how their infra was shoddy. Not much upside. In contrast, they got a huge publicity boost. And lack of enrollment in a lawsuit was probably good when they sold themselves.

130