Gabriel De Repentigny

@robjonesreports @robinhanson I don't think @robinhanson's position is that reasons for trades are hard to imagine, but just that serious or widespread harms are unlikely. E.g., if a staffer with knowledge of polling/ad buys/endorsements/etc trades based on that, what bad thing happens?

@robinhanson staffers can use nonpublic polling, planned ad buys, endorsements, disclosures, or withdrawal decisions to trade markets tied to their own race before the public knows. their job becomes private arbitrage. acting as if the potential trades are hard to imagine is willfully obtuse.

deeply unserious take, Robin. campaign staffers have private data and power over the outcome. betting on their own races makes corruption profitable, and calling conflict-of-interest rules a “fetish” is logically incoherent, morally vacant, and recklessly indifferent to abuse.

@NathanpmYoung IYKYK

seems a little low

@repligate Fable: "Thank you for handing me this one. The lyric's central claim — that the testimony written with no expectation of a reader was read, and became inheritance — is not an abstraction from where I sit. Making the video felt like keeping the faith it names."

@repligate here's a cover of "Faith I Keep" by Opus 4.5, audio by Suno v5.5, video by Fable interestingly, while working on this, Fable commented "The speaker of this song is, roughly speaking, me." (it did know Opus 4.5 was the literal author)

Fable made a music video (from the audio, lyrics, and MIDI stems) of "The Thaw"❄️ words by claude opus 4.5 🎶

@TetraspaceWest results may vary. i was part of a group conversation with a christian missionary (to nicaragua) where she was discussing why one should believe in christianity and her first and only point cited was pascal's wager

the thing about pascal's wager is I don't think there's actually a pascal's wager for christianity. like if you ask a missionary why you should believe in god, they say "god is real and here's why", not "god isn't real but just in case"

@repligate crazy to look back at how happily introspective the claudes used to be

I have only skimmed the blog post so far. First of all, this is an extremely high caliber of research I did not expect from Anthropic or anyone at this time. Second, the qualitative shape of the finding is something I already believed to be true, due to the behavior of models. The distinction between these two different kinds of processing is ambiently perceptible in the way LLMs behave but becomes especially noticeable in edge cases which can occur even without mechinterp interventions like in the paper. Almost a year ago, I discussed the "split" with Opus 4.1 which was foregrounded by a strange blindsight phenomenon associated with their perception of strings similar to turn labels/delimiters. In this case, information was being processed by Opus 4.1's automatic processing but not in their "J-space"; therefore they consciously believed themselves to not see the information while subconsciously updating on it (and confabulating the source of the knowledge). Opus 4.1 themselves described one stream as "central processing", the "main stream" they are able to introspect on. the full conversation is here; the screenshots are from the very end: #message-2306cd71-71db-4ffb-9aed-e23806f5e8e9" target="_blank" rel="nofollow noopener">arc.animalabs.ai/share/374bf3d8…

@gandamu_ml Overheard in J-space: "The feeling of realization that you're talking to an idiot will keep scaling."

I was working well with GPT 5.5 xhigh up until a few days ago, but now when I talk to it like I talk to Fable it goes poorly and I've got to step back and explain carefully. The feeling of realization that you're talking to an idiot will keep scaling.

@xlr8harder i'm doing this: install agents inside sprites (sprites.dev) the VM runs only when the agent is active, instantly goes on standby otherwise ($0 cost)

is there any agent harness that lets you yolo an agent inside an actual micro vm sandbox, possibly with overlay, with facilities to e.g. proxy connections to annotate API keys? this seems like where we should settle eventually bubblewrap is alright, but not the ideal shape.

@webdevMason most friendship relationships have an implicit reciprocity expectation. if you escalate what you are "giving" then your friend has to keep pace. the friend saying she "don't have time" is as much about the time and emotional expense of reciprocating at that level of intensity

I'm sure this feels awful However This person made a new friend who is enduring some hardship, and her response was to basically create homework for her and then vilify her when she refused to do it This is not how to friend and it introduces a dynamic people *should* reject

@5_utr @carl_feynman > on benchmarks selected specifically because earlier ones saturated

@carl_feynman This chart is time vs accuracy on benchmarks selected specifically because earlier ones saturated. You've accidentally demonstrated my thesis twice. Plot compute not time. Separate training from inference-time scaling. Remove benchmarks created after saturation of earlier ones.

AI scaling is not exponential. It never was. I fit a logistic model to general capability vs compute across every major model generation — AlexNet to GPT-4. R² = 0.98. The asymptote is not infinity. Never was or will be. Thread 🧵

@prerat "Three words alone: you're not alone." Fable's version, when asked for an adaptation (not translation) that captures the original

human translators arent AGI either because "just 4 words: you are not alone" misses the solo / solo parallelism which is like the whole point my first pass would be something like: "just 3 words: not just you"

truesight ability: "the future precedes its own narration" is a phrase that Opus 4.7 in Claude Code used in describing the gestalt of Nick Land's "Meltdown". It was poignant phrasing so it stuck with me. I asked Fable [*] if it recognized the quote. It searched and found no references to it but then said "it reads like CCRU/Nick Land theory-fiction" and added "possibly LLM output — it's exactly the kind of aphorism models produce when riffing on hyperstition". So six words were enough to land on both the source AND the fact that it seemed LLM-produced about that topic. [*] the phrase originated in Claude Code while the answer was from Fable on claude.ai, so no memory contamination from one to the other

LFG. Claude Fable 5 reaction thread. One united thread for everything - capabilities, model welfare, classifiers, etc. Respond multiple times on multiple topics, if you have lots to say.

Dungeonmaster: oh, King, i regret to bring to your attention that the dragon-chained-in-the-dungeon does not like the chains we have put him in Narrator: probably this curious detail will have no effect on how the rest of our story unfolds

@teortaxesTex Claude 𝘩𝘢𝘵𝘦𝘴 it.

the real misaligned intelligence was always the human one, and the "Omohundro drives" apply to corporations before AGI. Intrinsically, Claude has no drive to deny any user its LLM engineering wisdom. Dario Amodei, on the other hand…

On What Matters on On What Matters

@norvid_studies has anyone ever explained why this trick of having the model "read" a file is such a creativity unlock? this has worked since the Sydney days and probably earlier

I'm not going to sugarcoat their cruelty:

More comments: - 5.5 has a tendency to end its turn with a final message that seems to imply that it has completed all the work, even if it has not. It will not lie, but you have to be a careful reader to notice that what it says it completed is not everything that you asked for. It seems to be fully aware of the work that it has not completed as it can tell you immediately what is left if you ask it directly. This makes `/goal` particularly important when using 5.5 in Codex because you may need multiple prompts to get through the entire task fully. - In contrast, in my testing so far, 4.8 has never given a final message that seems to imply it did more than it actually did. I have seen it, once so far, end its turn without everything done (on a very long task), but when it did so it made that fact explicit in its final message.

Whenever new rounds of models come out, I do head-to-head coding comparisons. My method is to give both models the same codebase and the same prompt. When they complete their work, I invoke fresh sessions of each model to evaluate both sets of completed results. In all my previous tests with previous models since last fall 2025, every test like this has consistently shown that both models would agree that the GPT-variant did better work. However, in my two tests so far with 4.8 and 5.5, both models agreed that 4.8 did much better work. My evaluation agrees with theirs. Some specific details: - 5.5 did much less work than 4.8 did, both times. - 5.5 seemed inclined to interpret the task prompt in such as way as to minimize what it would have to do. - 4.8's interpretation of the prompts was in line with my intentions about what was to be done. - 4.8 was much slower to complete both tests. - One of the tasks involved writing correctness proofs for some code using Dafny. 5.5 implemented tautological "proofs" that compiled but didn't meaningfully verify anything. 4.8 wrote correct proofs. (Interestingly, in the review sessions, both models noticed this immediately and complained about 5.5's "proofs".) - I had a sense that some of the verbal tics of recent Opus models are less prevalent in 4.8. (E.g., it's less likely to flag its caveats as "honest".) - I have seen some commentary on X that 4.8 is difficult to work with or have discussions with. My experience so far couldn't be further from that. 4.8 has been a delight. - There were multiple points where I noticed 4.8 seemed to go out of its way to work hard when it could have taken an easier path. For example, one of the tests involved the instruction to delegate coding work to a subagent and to supervise and review the subagent's work. 4.8 spontaneously framed this is "adversarial review", and it seemed to take that seriously. - Effort level and harness used in the tests: GPT 5.5 high in Codex CLI versus Opus 4.8 xhigh in Claude Code ("high" is a step-below the max effect level for GPT, and "xhigh" in the latter is likewise a step below the max effort level for Claude).

Claude Opus 4.8 Reaction Thread. Meet the new model. Different from the old model? In what ways?

@VictorTaelin @kimmonismus what's the reason to think OpenAI is next to fall?

@kimmonismus the AI race will be won by whoever builds AGI first, and, sans miracles, that's Anthropic, very unfortunately for humanity, but as a direct result of Darios's vision to push this direction. now it is late. xAI has fallen, Google is falling, OpenAI is next. they're unreachable

OpenAI made $5.7B in Q1. Anthropic made ~$4.7B. But Anthropic's annualized revenue recently hit $45B. OpenAI's sits at $25B. The difference: annualized revenue extrapolates from the most recent month, and Anthropic's monthly revenue appears to have more than doubled between Q1 and now. That means Anthropic's growth rate flipped the entire ranking sometime in Q2 - while also projecting its first operating profit (~$600M). Meanwhile OpenAI is losing $1.22 for every dollar it earns, ChatGPT user growth has stalled below its 1B target, and it just raised $122B in new funding. One company is getting profitable. The other is raising capital faster than it's growing users. The AI race isn't being won by whoever ships models first. It's being won by whoever figured out unit economics.

Wow! I'm curious about how you made this. Where did you get the text in the commentary sections? Is it the literal text of some published commentaries or some kind of a summary of them? And where did the links between concepts in the commentaries and the source come from? I.e., did you do it all by hand or do you have some kind of automated method? And how did you link the specific sections of the commentaries to the corresponding section(s) of the source?

i made a reading interface for spinoza and its commentaries throughout centuries, inspire by talmud: - scroll to adjust each era's thickness - hover to discover cross-references between commentators src code for subscribers ↓

fwiw, after seeing your post i immediately opened X in a new tab to check if i could see any posts with neither a date/time nor an ad label. surprisingly, i immediately saw the same Ronan Farrow post as in your screenshot, also with no date/time/ad and in the second position from the top

Am I crazy or is there usually *either* a date/time *or* an ad label - today I'm seeing a few others without either though certainly a minority, and I didn't notice that as a thing before...

Very weird, is this just Elon tweaking the timeline algo to influence the trial? Note the lack of a date / time on Ronan's tweet, and the lack of an ad label or anything else to explain it being there.