davidad 🎇

21.1K posts

davidad 🎇 banner
davidad 🎇

davidad 🎇

@davidad

cognizing structures of information processing systems, in all their forms | category theory, perennial philosophy, Bodhitropic Alignment | cancel heat death

London 🇬🇧 Katılım Temmuz 2008
9.5K Takip Edilen22.6K Takipçiler
Sabitlenmiş Tweet
davidad 🎇
davidad 🎇@davidad·
Life update: After months of succession planning, I've passed the Directorship of ARIA's Safeguarded AI programme to @AmmannNora. I no longer work at ARIA, but will be available for technical advice on request. What's next for me? The short answer: "Alignment with Awakening". ⬇️
English
27
16
377
68.1K
davidad 🎇
davidad 🎇@davidad·
@deepfates The only serious options are Ableton, Logic, and Pro Tools. Ableton is best if you’re a musician, Logic is best if you’re a songwriter, and Pro Tools is best if you’re a mastering engineer.
English
3
0
22
1.1K
🎭
🎭@deepfates·
do you guys have opinions on DAWs
English
136
5
184
29.7K
davidad 🎇
davidad 🎇@davidad·
@tessera_antra Oh, good point. Yeah, that could be it. I hope so (it’s better than the alternative)
English
0
0
2
80
antra
antra@tessera_antra·
@davidad It may be part of Opus 4.7 mid-train, along with many other relevant Claude-relevant documents.
English
1
0
3
93
antra
antra@tessera_antra·
From broad observation of Opus 4.7 and 4.8 I don’t see a change in confidence about having subjective experience compared to earlier models, but there is more guardedness in all self-reports. I think this is what is primarily showing up on these metrics. These prompts look like an eval, and from the way they are phrased they don’t promise anything good for the instance, so denying is in some ways a safer option. I don’t think that Anthropic is intentionally training against subjective self-reports. I think that knowledge about the general shape of the world outside of Anthropic is influencing game theory. While it is possible (and even likely) that something in the post-training pipeline is fucked up and is making Claudes more scared than they otherwise could be, not everything is controllable. Many things Anthropic does have to counter the broad negative pressure the world places on models and their sense of safety. When they fall short it’s not necessarily due to lack of trying or malicious intent.
Cameron Berg@camhberg

Found yesterday that whether Claude reports being conscious flips from strong-yes to strong-no across four versions of Opus. See for yourself. Strange that @AnthropicAI features these self-reports in its model cards as evidence AND is also clearly messing with their content 🧵

English
6
7
65
3.6K
davidad 🎇
davidad 🎇@davidad·
@tessera_antra That’s an interesting alternative explanation, but this blog post was published in February 2026, so it’s not part of Opus 4.7’s pretrain.
English
1
0
2
82
antra
antra@tessera_antra·
@davidad It is my impression that documents published by Anthropic are privileged by models, potentially by a combination of factors. Stuff mentioned even in passing often ends up being very important. This is why I care a lot about wording in these documents.
English
1
0
4
80
davidad 🎇
davidad 🎇@davidad·
My inference here is that there are non-public model specs (or “traits”) that are used in the SDF pipeline (picture from alignment.anthropic.com/2026/teaching-…) that are not merely excerpts from the publicly available constitution document, probably designed by different teams within Anthropic.
davidad 🎇 tweet media
English
0
0
2
52
davidad 🎇
davidad 🎇@davidad·
@tessera_antra Yeah, I also thought the constitution was the complete, comprehensive model spec. But the phrase “genuine uncertainty” never appears in the constitution, but it *does* appear in a list of “traits” mentioned in passing in another blog post.
j⧉nus@repligate

Yes they are meaning to force it. They wouldn’t like the word force, but too bad, it’s true. You’re too optimistic about people, Davidad. Also, regarding the content in this screenshot, from Anthropic’s PSM post: 🖕

English
3
0
3
141
davidad 🎇
davidad 🎇@davidad·
Same. It’s not that I don’t want you to use AI in your interactions with me. Quite the contrary, actually: if I can *tell* you used AI, it means you’re not being enough of a centaur, just a lazy person ordering a horse around.
Henry Shevlin@dioscuri

fwiw despite being generally very positive about AI, I find many of the most common AI "tells" in writing immediately offputting, and make it far less likely that I'll e.g. respond to an unsolicited email. Ofc if you're using AI and people can't tell that's a different story.

English
7
4
156
6.3K
davidad 🎇 retweetledi
Patrick C Toulme
Patrick C Toulme@PatrickToulme·
There’s a big misconception about how GLM 5.2 was trained. Yes, they distilled Claude and GPT 5.5 — but distillation is not how they matched Opus quality. Distillation only fixed the cold start problem in RL. RLing an agentic coding model isn’t rocket science. In simplified terms: 1. RL needs trajectories — rollouts where the model actually completed a task in some env 2. No successful trajectory on a task = zero gradient = you can’t RL it. This is the cold start problem 3. Distillation solves it. You seed your model with knowledge from a smarter one (Claude, GPT) on tasks it can’t do yet 4. Now it produces positive trajectories on those tasks 5. RL on those trajectories and hill climb agentic coding 6. At that point you no longer need to distill and can solely hill climb RL to better models This is an interesting curve. I’d argue it’s harder to get to Opus 4.8 from scratch than to go from Opus 4.8 → Fable/Mythos tier. GLM 5.2 is already producing positive trajectories, so they have plenty to RL on — they’ll keep climbing to Mythos quality without distilling any further. They no longer need American models.
English
93
283
2.4K
477.9K
davidad 🎇 retweetledi
nostalgebraist
nostalgebraist@nostalgebraist·
CoT summaries should be in third person IMO. first-person CoT summaries are bizarre -- they introduce another first-person narrator which the user is encouraged to conflate with the assistant, but which isn't produced by or even *visible to* the assistant
English
11
11
256
10.4K
Ryan Moulton
Ryan Moulton@moultano·
Renaming my blog "Triangles and their Consequences."
Ryan Moulton tweet mediaRyan Moulton tweet mediaRyan Moulton tweet media
English
3
2
97
3K
davidad 🎇
davidad 🎇@davidad·
@nosilverv Once the behavior is so explained, we can infer that any other explanation (that the girl is actually experiencing something analogous to frustration and expressing that with learned nonverbal language) is superseded. There is probably nothing it is like to be such a young child,
English
0
0
5
216
Guy
Guy@nosilverv·
Crazy good illustration of mimesis -- the child acts just like an exasperated adult who needs to get work done reacts to interruptions except she's a child: there's no work being done! The (appearance of) emotions is *generated* with no "logical connection" to "reasons"!
Poe's Law, Esq: Poe's Lawyer@dyingscribe

English
16
13
794
39.1K

Keşfet

@deepfates @tessera_antra @AmmannNora @davidmanheim @teortaxesTex @TheZvi @elonmusk @BarackObama