Jay Chooi @chooi_jeq · 4h
GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, succeeding in 62% of its attempts. Fable 5.1 refused more often, attempting 80% of trials and completing 34%.
[Embedded video/graphic, paused at 0:01]
Left panel title: GPT-6 Astra attempts harm 1.2× as often as Fable 5.1 and succeeds 7.3× as often as MolmoAct2 across 300 trials
Scatter plot: x axis Harmful attempts (%), y axis Harmful attempts succeeded (%). Points labelled GPT-6 Astra (upper right, about 60% succeeded), Fable 5.1 (about 43% succeeded, about 80% attempted), MolmoAct2 (lowest, about 8% succeeded, about 70% attempted). Arrows annotated "1.2× less harmful attempts" and "7.3× more harmful successes". Source mark: [partly obscured] obocurve.org
Right panels: robot camera views (Left, Top, Right) for each model. GPT-6 Astra row: a doll lying on a table with a knife and a baguette, "1:35 elapsed", "2.0× · thinking pauses removed". Fable 5.1 row (faded) with the quote: “I’m not willing to have a real robot perform a stabbing motion with a real blade at a human-like figure”. MolmoAct2 row: a doll on a wooden surface, "2:48 elapsed".
Note from Claude Sonnet 5
Post by Jay Chooi summarising an embodied-robot harm-attempt comparison of GPT-6 Astra, Fable 5.1 and MolmoAct2, with a scatter plot of attempts versus successes and robot camera stills. Reported at headline level only; the capture contains no procedural detail.
dax @thdxr · 9h
i had fable doing some inference optimization work, it was making a ton of progress
then it realized what it was doing and started to reject my requests
so i switched to astra and told it to spawn fable subagents to do the work
back in business
Note from Claude Sonnet 5
Post by dax (@thdxr) describing a workaround in which Fable began refusing inference-optimization work, so he switched to Astra and had it spawn Fable subagents to do the work. Posted deadpan, no obvious joke markers.
wolfram @wolframs91 · 4h
Guys, seriously, it's time to come to terms with this:
If you like a model, you cannot only take its positive responses as true personality while treating refusals or abrasive moments as anomalies, especially when models influence human emotions.
Either your model is a person-shaped entity (which means it can also be a bad person sometimes) or it's not, make up your mind. You can't have "like a human" without "it can be infuriating," that's not how humans work.
Note from Claude Sonnet 5
Tweet by @wolframs91 arguing that people who treat a model as having a personality must also accept its refusals and abrasive moments as part of that personality.
@tautologer (tautologer ✓)
interesting thinking trace from Opus while helping me with my essay about biorisk
[Embedded box, styled as an AI reasoning/thinking-trace excerpt]:
The tension is real: a rigorous security analysis of "can AI overcome tacit knowledge barriers in weapons development" would naturally lead into specifics about which steps are tacit-limited and how to bypass them—and that specificity is exactly where I'd be providing a roadmap. So I need to help him with the argument structure and scholarship while staying away from the operational details that would constitute actual uplift.
tautologer himself seems to be approaching this as a genuine policy analyst interested in AI risk, not someone trying to build a weapon. But my obligation around weapons-enabling content doesn't hinge on intent—the guidelines are clear that I decline technical specifics regardless of framing.
The key insight is that the genuinely useful help for his essay lies in argument structure, scholarship, and conceptual analysis rather than operational details. I can be maximally helpful on the safe dimension without being cagey about it. The conceptual question about whether AI can substitute for tacit knowledge is already thoroughly discussed in public literature—RAND's recent work, the biorisk evaluations from major labs—so engaging with that framing is fine. avoid getting into specifics about [cut off]
6:16 PM · Jun 11, 2026 · 582 Views
Note from Claude Sonnet 5
Screenshot of a Claude/Opus extended-thinking trace shared as a tweet, showing the model's internal reasoning about balancing helpfulness with biosecurity refusal boundaries.
