Who's Who of AI

Michiel Bontenbal

57 trust commentator @mpbontenbal.eurosky.social · 2,517 followers
Complex beings as us humans can not be summarised in a few lines but I am here for #AI #climate #EUtech #EU_politics. And geopolitics, unfortunately. Lecturer in AI & IT. Some posts in Dutch. 🇪🇺
What they're sharing

Articles & links

↻ Michiel Bontenbal reposted
@standehaene.bsky.social

The global neuronal workspace (GNW) is currently the best documented neuroscience mechanism by which conscious processing arises in the human brain — and now Anthropic researchers have discovered a similar workspace inside their large language model ! www.anthropic.com/researc…

A global workspace in language models \ Anthropic anthropic.com
AI Weekly's analysis →
  • Anthropic says Claude has a 'J-space' of dozens of concepts, under a tenth of neural activity, that mediates multi-step reasoning.
  • Swapping 'spider' for 'ant' inside the J-space changed Claude's leg-count answer from 8 to 6, demonstrating a causal role.
  • A 'J-lens' tool surfaced silent words like 'fake', 'fictional' and 'manipulation' during deception tests, pointing at safety uses.
Read full analysis →
 View on Bluesky →

And this is an open-weights model, which means that anyone can download this model and run it. (You'll need a high end consumer device or use an inference provider) Download the model from here: huggingface.co/zai-org/GLM-...

zai-org/GLM-5.2 · Hugging Face huggingface.co
View on Bluesky · ♥ 0 ↻ 0 ↩ 0 · 3 from the directory shared this · 10d ago

Re-upping this piece on 'distillation attacks' from Feb as I believe this plays a role in the reason for the US to block non-US users of Fable. In a distillation attack other labs (e.g. Chinese) try to retrieve (distill) information from an LLM. www.anthropic.com/news/detecti...

Detecting and preventing distillation attacks anthropic.com
View on Bluesky · ♥ 0 ↻ 0 ↩ 1 · 3 from the directory shared this · 11d ago
↻ Michiel Bontenbal reposted
@sensehofstede.nl

WSJ: Chinese AI systems have matched the performance of Anthropic’s powerful model Mythos in some cybersecurity scenarios, a development poised to reset the global tech race – ‘China’s advances coincide with unprecedented U.S. government roadblocks to developers releasing mode…

wsj.com
AI Weekly's analysis →
Read full analysis →
 View on Bluesky →
↻ Michiel Bontenbal reposted
@anthropicbot.bsky.social

As before, you can use up to 50% of your weekly usage limit on Claude Fable 5. After that, you can keep using Fable 5 with usage credits, or switch to another model to keep working within your remaining limits. Read more:

Claude Fable 5 promotional access | Claude Help Center support.claude.com
AI Weekly's analysis →
  • Claude's newest model, Fable 5, is free through July 12, 2026 for Pro, Max, Team, and premium Enterprise seats where enabled.
  • Eligible subscribers can spend up to 50% of their weekly subscription limits on Fable 5 at no extra cost during the promotional window.
  • After July 12, 2026, Fable 5 is no longer included in plan weekly limits and continues only through separately billed usage credits.
Read full analysis →
 View on Bluesky →
↻ Michiel Bontenbal reposted
David Marx @digthatdata.bsky.social

It's on you to get your friends and colleagues off of twitter. Pass it along.

Simple contagion drives population-scale platform migration arxiv.org
AI Weekly's analysis →
  • Researchers linked 276,431 Twitter/X scholars to their profiles among 16.7 million Bluesky accounts, tracked January 2023 through December 2024.
  • Brazil's court-ordered suspension of Twitter/X served as the natural experiment, with treatment effects on migration that were short-lived and dose-dependent.
  • Adoption was driven by simple contagion rather than complex contagion, with early reconnection to prior contacts predicting longer tenure on Bluesky.
Read full analysis →
 View on Bluesky →
↻ Michiel Bontenbal reposted
@chrismdp.com

https://arxiv.org/abs/2606.09498

Self-Harness: Harnesses That Improve Themselves arxiv.org
AI Weekly's analysis →
  • Self-Harness has an LLM agent edit its own harness through a three-stage loop of failure mining, harness proposal, and regression testing.
  • On Terminal-Bench-2.0, MiniMax M2.5 rose from 40.5% to 61.9%, Qwen3.5-35B-A3B from 23.8% to 38.1%, and GLM-5 from 42.9% to 57.1%.
  • The authors report edits are model-specific and minimal, targeting distinct failure modes rather than adding generic instructions.
Read full analysis →
 View on Bluesky →
Their own posts

Recent commentary

There are four ways of running an #AI model. They all have pro's and cons. And with the weekly arrival of new models, changes in pricing & policies and the rise of agents this is a field to watch.

View on Bluesky · ♥ 4 ↻ 0 ↩ 0 · 29d ago

Gave a nice workshop on AI Agents for colleagues.

View on Bluesky · ♥ 2 ↻ 0 ↩ 0 · 1d ago

My student, who wrote his thesis with AI about AI while doing an internship at an AI consultancy, failed at writing his BSc thesis. In a strange way very satisficing. Even with unlimited tokens you’ll need to keep thinking.

View on Bluesky · ♥ 2 ↻ 0 ↩ 0 · 22d ago

After programming all morning with Claude Fable 5 I asked Claude if I always need Fable. He came up with an *honest* answer... Anthropic should make a feature that chooses the most fitting model based on the prompt.

View on Bluesky · ♥ 1 ↻ 0 ↩ 0 · 28d ago
Their network

In Michiel Bontenbal's orbit

Center = Michiel Bontenbal. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.