ayewo

> Swing.invokeLater(closure)

Should be:

   SwingUtilities.invokeLater(closure)

Interesting post!

When you say "Full-time, 32h, 24h possible | €5,000/month (full-time)" is the €5,000/month salary pegged at 40h/week so that the actual payment is scaled up/down based on the number of hours worked per week?

> So long as I can rent an RTX PRO 4000 Blackwell at about half a buck per hour through a European provider, ..

Mind sharing the provider’s URL?

Or even better: use a (visiting) tourist.

> Fun fact: it's a crime to say you're a lawyer if you aren't one.

This is not an absolute fact.

It depends entirely on the commenter's jurisdiction.

Sounds similar to Ramp's Latent Briefing for multi-agent coordination.

https://x.com/RampLabs/status/2042672773747589588

It's impressive that you didn't reach for Claude Code first :)

Would be good to keep track of how long it takes you to complete, so you can compare it with how long it takes Opus and then Fable to accomplish the same thing.

To add to this, merely using the thumbs up/down button in a chat could share your entire conversation with them for model training.

From their docs[1] (archive copy is at [2]):

> You can opt out of training through our privacy portal by clicking on “do not train on my content.” To turn off training for your ChatGPT conversations and Codex tasks, follow the instructions in our Data Controls FAQ. Once you opt out, new conversations will not be used to train our models.

> For a linked teen account, a parent or guardian may manage whether conversations can be used to improve our models through Parental controls.

> Even if you have opted out of training, you can still choose to provide feedback to us about your interactions with our products (for instance, by selecting thumbs up or thumbs down on a model response). If you choose to provide feedback, the entire conversation associated with that feedback may be used to train our models.

[1] https://help.openai.com/en/articles/5722486-how-your-data-is...

[2] https://web.archive.org/web/20260910151242/https://help.open...

> Microsoft famously had teams at odds with each other, ...

Instantly reminded me of this especially apt comic:

https://pbs.twimg.com/media/EyJfRwHWYAI0_yt?format=jpg

[1] Also available here: https://www.reddit.com/r/ProgrammerHumor/comments/6jw33z/int...

>> Are we forgetting how fast they rushed in to deploy Claude at the DoW (Department of War) with Palantir?

> I hardly see how the Dow Jones in relevant here, that’s finance

Not Dow Jones. DoW = Department of War.

Not so sure since Anthropic has 4 model families while OpenAI has 3 for GPT-5.6.

Claude Fable/Mythos vs GPT-5.6 Sol

Claude Opus vs GPT-5.6 Terra

Claude Sonnet vs GPT-5.6 Luna

Claude Haiku vs ?

Spot on wrt CoT. I have thinkingSummaries enabled and I find it eminently readable compared to the prose in Claude's replies.

In fact, whenever Claude disobeys me, I usually first skim the CoT to figure out if my original instruction was ambigous given the context. I usually come away with a better understanding of how to frame my prompt to be less ambiguous or just force myself to be more explicit when prompting.

Regarding diosbedience, usually this is either due to a blanket instruction from me during an earlier turn in the same session, an explicit instruction in its system prompt or it being just eager to bring a task to completion.

  # ~/.claude/settings.json
  {
    "model": "opus",
    "showThinkingSummaries": true,
    "skipDangerousModePermissionPrompt": true,
    "verbose": true,
    "remoteControlAtStartup": true,
    "agentPushNotifEnabled": true
  }

True. But it was also meant as a counter to “Sent from my Blackberry”.

Obama and a lot of execs were pretty addicted to their BB back in those days.

Perhaps these may be Huawei Ascend chips.

https://en.wikipedia.org/wiki/HiSilicon#Ascend_910

https://medium.com/@huaweiclouddevelper/a-brief-introduction...

They want to be able to zoom in and zoom out as needed while analyzing aggregated user requests to better understand the different kinds of ways (well-resourced) actors use to distill their most capable models.

> "The data will help us defend against complex and novel attacks (including new jailbreaks and attacks that operate across many requests) as well as help us identify and reduce false positives."

From: https://www.anthropic.com/news/claude-fable-5-mythos-5

> "Some attacks only become visible across multiple requests. Best-of-N jailbreaking, for example, sends hundreds of slight variations of a prompt in the hope that one will work. Larger patterns of misuse, such as state-sponsored espionage or data extortion campaigns, only surface when our safeguards classifiers can zoom out across many requests. Detecting these threats requires temporarily retaining prompts and outputs so they can be analyzed together, rather than one at a time."

From: https://support.claude.com/en/articles/15425996-data-retenti...

Yup.

The original source was Kyle Daigle, GH COO.

- https://x.com/kdaigle/status/2040164759836778878

- https://xcancel.com/kdaigle/status/2040164759836778878

Isn’t that the optimal strategy?

That an LLM trained to be a paper-clip maximizer chose the optimal strategy is in my opinion the most plausible outcome.

> 1) AI researchers talk and change companies often, so techniques circulate. This feels implausible because training and shipping a new model ought to take longer than 2 months?

The assumed timeline (2 months) is slightly wrong because Fable (Latin) is essentially the same as Mythos (Greek) albeit with protections against cyber and biological misuse.

Mythos (Preview) was publicly announced in April 2026 [1] which means other labs have had 4 months to catch up, not 2 months.

Assuming everyone had access to Mythos from the start, your expression, similar to other folks would have been "Mythos-level intelligence" and not "Fable-level intelligence".

1: https://news.ycombinator.com/item?id=47679258

Thanks for the PyTorch internals OG link and for putting together your PyTorch one-pager, if we can call it that :)

Kudos.

Off-topic:

Some of your comments in this thread are dead, but I’ve vouched for some them as I don’t see anything wrong with what you’ve written, other than your writing here on HN sounds a little like you are using an LLM to compose each reply (or perhaps your regular use of LLMs is showing through in your comments?).

Perhaps this was due to their red-teaming partnership [1][2] with Anthropic which they wrote about a few months earlier in March?

1: https://www.anthropic.com/news/mozilla-firefox-security

2: https://blog.mozilla.org/en/firefox/hardening-firefox-anthro...

Previous discussion: https://news.ycombinator.com/item?id=47273854

How did you unearth such a remarkable find of 2 different faculty individuals named "Philipp Otto" collaborating on the same paper :) ?

Yep. Cursor’s Composer 2 model is a good example, though it is not clear if they entered into an agreement with Moonshot before they got found out in March this year [1] or after.

1: https://x.com/fynnso/status/2034706304875602030

The top poster mentioned LLM spend of millions/month to justify the estimated capex of $6m to self-host Kimi on own infra.

Add to this number another $1.5m/yr in opex, so not sure I’d call such an enterprise wealthy enough to spend those kinds of sums on LLMs an “SMB”.

Understood but sharing your existing devops resources with this will soon become a bottleneck especially when any major downtime will keep several engineers (and long-running agents) blocked from any meaningful work until availability improves.

They do. It’s a reasonable stance [1] that is aptly captured by this quote:

“I doubt very much if it is possible to teach anyone to understand anything, that is to say, to see how various parts of it relate to all the other parts, to have a model of the structure in one’s mind. We can give other people names, and lists, but we cannot give them our mental structures; they must build their own.” — John Holt

1: https://www.recurse.com/blog/191-developing-our-position-on-...

In your opinion, do you think Internet Protocol Version 8 (IPv8) [1] stands a chance to fix the mistakes of IPv6 after more than 20 years now?

Or there is too much inertia for IPv8 to overcome to become a truly backwards compatible extension / superset of IPv4?

Part of the reasons for the slow adoption of IPv6 was that it was never designed to be backwards compatible unlike IPv8.

1: https://www.ietf.org/archive/id/draft-thain-ipv8-00.html

I'm not sure if you are aware, but you have to approach prompting Fable slightly differently from a model like Opus.

It's important to include the reason aka the why of your task [1] in your prompt. You'll get more mileage if you verbalize your thought process when prompting Fable. Anthropic say you should think of Fable as a "thought partner".

1: https://platform.claude.com/docs/en/build-with-claude/prompt...

2: You might find some of the example prompts listed here useful https://x.com/trq212/status/2073100352921215386

> Sol had all the trimmings, Terra had the least.

That’s interesting. My understanding is that:

Sol = Sun Terra = Earth Luna = Moon

So it’s a bit surprising that in Toyota’s nomenclature, Terra is the basic trim instead of Luna.

A Google DeepMind researcher (Neel Nanda) was able to replicate their claims on an open weight model (Qwen 3.6 27B):

> We have replicated the core claims on Qwen 3.6 27B, and also share preliminary evidence of extending this work by finding abstract "interpretative meta-tokens", like Chinese characters for "what does this mean" that seem to activate and play a causal role on processing ambiguous sentences

See p33 of [1]

Anthropic also released companion code to go with their paper in [2] which also used Qwen. They state that their code should be broadly adaptable to other open weight models with HuggingFace decoders.

[1]: https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be24...

[2]: https://github.com/anthropics/jacobian-lens