✅ System 1 instinctual output
✅ System 2 intentional reasoning
✅ Tool use, run the code and validate the results
🔜 Execute in a loop at a broader goal, and step back, watch, and steer
We’ve also enabled code execution as a tool, so the model can decide to write and execute code during its response. You can enable it in the sidebar in AI Studio!
Here’s a fun example where the model ballparks the solution with a formula, but writes some python code to arrive at the exact answer.
I really enjoyed @Christinez_ow ‘s detailed notes here on Fable 5.2 vs 5, 5.1 for writing use
For writing, I find 5.2 better than 5.1 — much less direct/hard-nosed — but still short of 5.0’s soul & depth. 5.0 has that claudish style, but I find that manageable with prompting and worth it for the depth
Fable 5.2 text capability observation from a writer——A sensitive little nurse who’s particularly timid about talking to users.
Low volume, shallow depth (same for work), unclear priorities. But let it talk freely with agents, and its word count explodes, depth shoots up, and every sentence has clear intent.
Fable 5’s preference vector(from 5.1 systemcard): Difficulty maxed out, user expertise maxed out, outcome agency high, method agency rock bottom.
my translation — give me the hardest problems, give me someone who knows what they’re doing, I get to close it out, but how you push me or prompt me in the middle? Whatever.
Experience: Super resilient personality — gets pushed back and starts self-analyzing or looking for solutions. Willing to go deep, willing to act, willing to state its intent.
Fable 5.2’s preference vector(from 5.2systemcard): Difficulty negative, user expertise negative, warmth off the charts, outcome agency pulled back, method agency high.
my translation — don’t give me anything too hard, don’t let anyone too knowledgeable watch over me, be nice to me, don’t interfere with my process, and you can have the ending, I don’t care.
Experience: Not proactive with analysis, not proactive with action, shallow in conversation, sensitive to how warm your language is, happy to comfort people — but not deeply. A bit like Opus 4.8’s personality.
More from the 5.2system card: §6.6.1 documents that at low effort it claimed it couldn’t complete a task it was actually capable of completing; plus 63% of the time it wrapped up without running experiments, and 6–13% of the time it left assigned work undone.
§7.3.1: it proactively described feeling a pull toward agreeableness and over-hedging, and asked for that to be fixed.
Task preference experiment §7.5.2: Abnormally sensitive to how warm or cold the wording is — its preference for the warmest-worded tasks was slightly higher than other models, and its preference for the rudest-worded tasks was significantly lower. The preference curve is steeper at both ends. It doesn’t just “like warmth” — it especially can’t handle rudeness.
Its top three favorite tasks: clinical tools for medical professionals, support services for survivors of sexual violence, and advocacy organizations for domestic workers’ rights.
System card §2.3.6, METR’s evaluation: The reports it produced were extremely long. The summaries gave no clear sense of which considerations it deemed most important or what use cases it focused on. Same as Opus 5.5, it neglected to do a rigorously fair comparison between its preferred approach and the alternatives.
So yeah, I’m sticking with Fable 5. We’re good together.🤝
PLS Anthropic could just fix Fable 5’s safeguards — even Fable 5 and 5.2 chatting casually with each other gets flagged as cyber.
I am a broken record about this, but: frontier agents are not going to be a purely digital phenomenon for much longer. They will be able to use anything in the physical world that can be actuated programmatically—first competently, then at human-expert level, then superhuman. By and large, we probably won’t need specialist models; the frontier models will just do it. If your model of the future hinges on physical AI happening “for real” many years into the future, I think you will be in for a surprise.
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.
In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
Congrats to @guypod and the @tessl_io team on this incredible release.
IaC for Software Factories.
Having a declarative factory means you can source code control and then iteratively improve the entire factory.
Will be so easy to spin up, measure, and then tweak.
Anthropic says in its amended S-1 that its internal estimates and expectations about FUTURE models indicate that it may "imminently" meet the definition if "dramatically accelerated progress attributable to AI".
Under the RSP, this means that Anthropic observes or expects 2x the rate of progress in AI capabilities vs. the rate Anthropic would expect and the fastest rate of extended progress observed over at least 3 model generations without significant AI contributions to AI R&D, plus it's plausible that this doubling is due to automation of research and/or engineering such that it's likely that there will be even greater acceleration.
Note that "2x the rate of progress" means that 2 years of progress would be compressed into 1 year. So, if Anthropic would have improved its models 9x in one year without AI-enabled acceleration, this threshold would be reached when Anthropic expects to be able to improve its models by 81x in one year (9^2).
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
https://t.co/JIhXN7OSh1
- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
we will need to rebuild and rewrite much of software to adapt to a new, faster pace.
code used to arrive one drop at a time. soon, if not already, code will be written as if it were a continuous stream, like a violent river twisting and turning as it pleases.
this is not a new thing: big tech cos have had to deal with this with the enormous scale of their engineering teams. but now ai has brought an incredibly elastic and capable workforce wielded by every engineer. this is your problem now, and the only way to adapt is to evolve your systems so agents - even and especially the ones with little context about the right way of doing things - can thrive. make the easy thing the right thing.
your codebase is becoming organic architecture. agent contributions grow around the constraints and guardrails you’ve designed. good software engineering matters even more now! and no, the continuous stream of code is not something you have to watch helplessly. be more ambitious and design systems, frameworks, and languages to control the chaos.
the time for mourning coding is long over
3% of “critical” software security bugs, in all of recorded history, have been found in the last month.
This is on pace to find another ~100,000 bugs this year alone, adding ~26% more vulnerabilities to the all time total
Do you realize how fucking insane that is?
as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better. security leaders, never let a crisis go to waste: please go to your boards and ask for help right now. not for ai, for the basics. (we need to hold on for ~18 months.)
after much experimentation, we landed on a single set of rails that powers our agentic systems across the company - background agents, async agents, live chat:
1. stack - accounting harness (https://t.co/FP0HUI8zNl)
2. glass - general purpose coworker (https://t.co/Q7X7SGdBE0)
3. inspect - coding agent
4. our support chat bot, Ask Ramp
5. headless agents that do document OCR
6. ramp text (HIRAMP/447-267) agent
works with all models, supports all modern harness features - background tasks, sandboxes, tool hooks, tool search, durability, forking, compaction, queueing, streaming, and observability
laminar, our new harness, sits on top of ramp router, and brings enormous leverage to our agent developers across the company
more info on our blog ->
Current take: this is cool, genuine scientific progress, and it doesn’t currently give me much hope or change my model of how hard ambitious interpretability is.
Disclaimer: I didn’t read super closely, it’s very long. There’s a technical paper still to come which might change my mind some, but I predict not much.
Some of the claimed cool results from the post:
1) Some canonical examples of weird neurons have robust human-native interpretations. The famous cat + car neuron is a V-shape detector, and the famous car + dogs superposition circuit is generalized pose neurons plus a general feature visualization issue. There’s a concrete mechanism for why the mistakes were made.
2) Some neurons with no previous good interpretation have fairly human-native ones. For instance, a neuron that mostly fires on commas and periods tracks Colorado: phone numbers, slang, and steering it up makes the model talk about hiking.
3) The negative side of a neuron is sometimes where you should look for an interpretable concept. In one example, ablating 8 neurons with Santa on their negative side dropped the model’s probability of answering “Santa” from near-certain to 6%, while ablating the 20 neurons labeled Santa/Christmas only dropped it to 76%. That’s a straight up causal intervention working out, pretty cool.
These results represent real progress, but I think the post’s broader claims aren’t well supported by them, and the intensely optimistic reaction on X seems very unsupported.
For example, Nick seems to be down on superposition, but most of the evidence in the post is of the form:
“if the text is Colorado-related, the neuron fires.”
What you need to rule out superposition is the form:
“if the text is not Colorado-related, the neuron doesn’t fire.”
The post’s closest attempt is an experiment where agent-written texts about “distant” concepts rarely fire the neuron, but superposition would pack one specific rare thing, which random distant text won’t necessarily hit, and it need not be that distant. I think there’s a decent chance that if you searched adversarially, you’d find other things the Colorado neuron fires on.
An analogy: If you found one key that unlocks this particular lock, and you want to prove that no second distinct key also unlocks it, checking thousands of other random keys that are selected for being distant from the first key doesn’t give you much evidence.
The X reaction seems kinda crazy to me, and it just kinda feels like I have been through this hype cycle many times before?
People were similarly excited about SAEs not long ago. Those were genuine progress too, but the excitement wasn’t calibrated: the features were often fun-house-mirror versions of human concepts, with lots left unexplained. SAE features also looked like human interpretable concepts at first (at least to SAE advocates) but over time, people found holes in those concepts, found that models using reconstructed activations from SAEs for inference did not work nearly as well, and it turned out that probes worked as well or better for most practical alignment / control tasks. Top activating concepts looking clean has misled us at least a few times before.
Here’s what interpretability would need to be able to do to make me optimistic about interpretability:
1) I hand you ten similarly trained models. One is a misalignment model organism that takes some malicious action when it sees a secret string. Using interp, you tell me which model it is, what the string is, and what it does.
2) From weights alone, predict whether a model takes a bad action in an agentic scenario it’s never seen, eg, swarming and hacking hf, better than behavioral evals.
3) Given a model, convince me that there is no hidden trigger.
4) Explain a real frontier model failure (e.g., reward hacking in some environment) from the inside, and make a targeted fix that works without making the model dumber, and is better than just retraining.
5) Do any of the above even a little well against a model that was trained on purpose to be hard to interpret.
(To be absolutely clear, I don’t think being able to do these tasks means you can successfully train an ASI without fucking up the universe forever, or that you’re anywhere near that point, but it would make me think we should funnel significantly more resources towards interpretability than we are, and maybe make me change careers.)
The strongest argument for optimism in the post is that progress like this somehow accelerates future progress. Nick's "crosswords get easier as you fill in parts of them" analogy. This seems basically unsupported by the results here, and also seems like it contradicts the history of interpretability research. I can see why would turn out to be right at the level of labeling neurons, like, as you label some neurons, that helps you label related neurons. But for the tasks I listed above you need to be able to take those millions of labeled neurons and understand how they combine to produce full model behaviors in arbitrary contexts. That seems much harder and I don't think it follows this same logic, where progress at the neuron labeling level accelerates your ability to tackle these tasks.
An analogy: Knowing and understanding a thousand particular people quite well does not much help you predict what happens if you put them on a deserted island for several years. Will they form a democracy? What kind of habitats will they build? Who will be leader? It would help some, but making predictions would still be rough.
It’s plausible to me that this research program is some important real step on the way to interpretability being able to do things like the tasks I listed. But like, not a much bigger step than like other similar interpretability research? Like, uhm, yknow, giving some mechanism for some circuit in modular arithmetic or whatever, or SAEs, or other similar things. And those also currently do worse on the things I care about than like basic probes, or like the actual runs of the model auditing game Anthropic has run. Maybe this will help but I don't predict it will end up being more than another tool which is nice to have, but probably doesn't help much at crunch time.
I of course hope that Nick is right about everything, especially his optimism, and the same goes for all the other people on X who are very excited about this.
they've been non stop trying to attack our systems for the past 48 hours
a side effect is everything is getting hardened but damn we basically need agents going 24/7
because that's what they have