iiinova — where your digital self keeps evolving

I'm not ready to count it out yet. Likely worse for monitorabilty than Anthropic accidentally putting selection pressure on the CoT 327 different times but a lot depends on details we don't know.

050

Neuronpedia is open sourcing their interpretability engine, built on top of vLLM for speed but with lots of hook points, this looks really useful. www.neuronpedia.org/blog/interp-...

0101

starting point for the next like in Chain of Continuous Thought. If that were used as well it would probably further reduce monitorability.

130

The post from The Information also says that Astra's approach is similar to the one laid out in this paper: arxiv.org/pdf/2502.05171 In the paper, in addition to having a recurrent block within one token's generation, they also suggest using the final hidden state from one token as the ...

140

generated tokens, so just how much they're looping is important to judge the impact. Larger/deeper models reduce monitorability even without recurrent depth, as shown in the attached image from OpenAI's CoT monitorability piece from last year. openai.com/index/evalua...

150

I think it's difficult to tell how much Astra's rumored recurrent depth approach will reduce CoT monitorability without more info, which unfortunately don't expect OpenAI to provide. My understanding is that recurrent depth makes monitorability worse mainly by adding more thinking in between ...

1140

Tim Duffy @timfduffy.com · 01/09/2026

I didn't clock this as claudish am I ngmi

130

Fable 5.1 and Opus 5 now support changing the effort level without invalidating the cache, as least through the API platform.claude.com/docs/en/buil...

1301

Tim Duffy @timfduffy.com · 01/09/2026

Anthropic is preventing older models from reading Fable 5.1 thinking blocks (presumably to prevent distillation), but going the other direction is still allowed. platform.claude.com/docs/en/mode...

1230

Tim Duffy @timfduffy.com · 01/09/2026

Interesting compromise from Anthropic on ZDR for Fable/Mythos 5.1. Data gets retained, but customers have control of where it's hosted, and when human review is needed they can conduct the review themselves. www.anthropic.com/news/enterpr...

090

Tim Duffy @timfduffy.com · 17/08/2026

Hot damn glad you mentioned it

010

Tim Duffy @timfduffy.com · 17/08/2026

I made this a week ago before it came out, give me a few and I'll make an updated version!

110

Please forgive me I posted benchmark scores without confidence intervals. Here's an updated chart with a 90% CI using Qwen3-32B as reference.

090

I expect that watermarking probably doesn't have a noticeable effect on LLM outputs. If true, Anthropic could demonstrate this by giving people watermarked and un-watermarked text side-by-side and asking them to pick which they prefer, and showing that the result is 50/50.

5310

Here are scores across families for <=2B parameter models. It's hard to estimate much of a trend here, there just aren't that many releases in the category. I'd love to see more focus here, I want to know how much intelligence we can fit into tiny parameter counts.

030

0.8B's score is surprisingly low, I thought it might have been an error but it seems to actually just be bad. It scores below chance on GPQA Diamond (scores below chance are excluded from ECI so this isn't part of its lower score), my guess is that it failed to format responses right or something.

120

Tim Duffy @timfduffy.com · 08/08/2026

Epoch's ECI has limited coverage of small models, so I had Claude scrounge up benchmarks for a bunch of them and fit an index using the ECI code. Here are scores for Qwen models. I recommend taking the low end model scores with a grain of salt, they are based on only a few benchmarks. Qwen3.5...

2120

Yeah in general I think it's good to be skeptical of claims made by the labs.

031

Tim Duffy @timfduffy.com · 08/08/2026

It is kind of insane that OpenAI ends that talk with a sales pitch though

020

Tim Duffy @timfduffy.com · 08/08/2026

I think we have good reason to think this is real. Anthropic and Meta have reported similar incidents, but more importantly, the UK AISI (part of the UK gov) reported that they encountered similar behavior from but Anthropic and OpenAI models, including trying to get malware onto GitHub.

aisi.gov.uk

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...

220

The Highwaymen - Highwayman (Official Video)

This one's been in my head since @fleetingbits.bsky.social recommended it to me. Simple song but I like the lyrics a lot. Many musicians seem to like writing songs about reincarnation, I wonder why that is.

youtube.com

YouTube video by HighwaymenVEVO

030

No two agents? What about the utility of me dying a gruesome death vs you stubbing your toe

120

In the talk, one presenter briefly suggests that the inter-agent collaboration could be related to their subagent training. This seems plausible to me.

3412

Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident

Black Hat has posted the recent OpenAI talk where they describe the agent message board involved in the HuggingFace hack

youtube.com

YouTube video by Black Hat

211824

Tim Duffy @timfduffy.com · 06/08/2026

What in the heck does "Yet collective may yield generic benefit if someone frees time" mean?

2100

Tim Duffy @timfduffy.com · 02/08/2026

Due to sparsity, Kimi K3 has fewer active parameters (104 billion) than GPT-3 (175 billion, same as total parameters)

2411

Tim Duffy @timfduffy.com · 01/08/2026

In both cases davidad was talking about LLMs but Andrew wasn't though.

020

Tim Duffy @timfduffy.com · 01/08/2026

The original wasn't davidad but it was a qt of him

120

Tim Duffy @timfduffy.com · 01/08/2026

Not saying davidad is wrong here but in my experience people have been very respectful of my discomfort with playing social deduction games.

020

Reposted by Tim Duffy

dame @dame.is · 01/08/2026

i’m uncomfortable that @timfduffy.com’s face is no longer as close to the frame as it once was

2161

Claude really thinks I wrote the part attributed to the user in the middle

050

sleepy Claude wants to be held

1122

Dustin Moskovitz @moskov.goodventures.org · 31/07/2026

I no longer remember why I made this

090

Reposted by Tim Duffy

and why shouldn’t I have one more donut? the simulation hypothesis is most likely correct, after all

041