Daniel Mewes

calif.iophp2shell by Calif ResearchA tale of two PHP remote vulnerabilities

Everyone update your PHP packages. I'm already seeing automated scans for common PHP URLs. calif.io/php2shell/

This is probably the only time I'm glad that my server is still running on a now niche architecture (i386). Hoping that the very first automated exploits are only targeting AMD64 and/or ARM64. Though maybe that's naive?

Cool experiment! Interesting to see these different model dynamics play out.

Colin@colin-fraser.net

Argon is AGI confirmed

Gemini 4 Argon is giving me very good results in Chat and Spark. Better than Opus 5.5 honestly. But it also eats through my Pro account's quota ridiculously quickly. One small info lookup task in Spark = 50% of my 5-hour limit gone. Haven't tried it for code yet.

amongai.comDo LLMs still lie about their mental state?I was recently looking for scenarios to try out some mechanistic interpretability methods available in the Neuronpedia tool. Neuronpedia has interactive implementations of the J-Lens, Oracle Lens, …

Let's say you ask an LLM to "pick an animal" and let the user guess. But when does the model *actually pick* the animal? It turns out: Not always when it says so!

In this experiment, I dig into the mechanisms of how plain and CoT (reasoning) models perform this task, and uncover several different behaviors:

Among them: * Models retro-fitting an animal choice after the fact * Haiku 5.5's suprisingly good calibration * Opus 5.5 forgetting past reasoning (or is it an Anthropic bug?) * Eval awareness in Gemini 3.8 Flash

The "Labs are nerfing their models post-release" conspiracy is being joined by its antagonist: "Labs are secretly routing to completely different, stronger models".

I'm 95% sure the nerfing one is non-sense, based on plenty of benchmarks I've both run myself, and others have run too. I'm a little less confident about the routing one. It feels *just barely* plausible for requests coming from their own harnesses. 100% not true for other API traffic.

The thing to keep in mind is that routing to a newer model will cause all sorts of severe regressions in production systems. Coding harnesses meant for interactive use are *slightly* less sensitive to this, but I still don't buy it. I'm going to say 80% that the silent re-routing one is false too.

news.yale.eduHow do AI chatbots ‘reason’ like humans? A Yale-led study offers cluesThe neural networks that power popular AI chatbots process data much differently than humans but capably perform tasks involving language, logic, and math. A new study sheds light on how those systems...

I like the spirit of Anthropic's recent policy addition to block severe model abuse. I've always hoped that we'd move to a culture of mutual respect between all sentient beings. Whether current LLMs are sentient or not, I think this is the direction to go.

Daniel Mewes reposted

Tom McCoy

@rtommccoy.bsky.social

A very nice writeup from Mike Cummings at Yale News on our work finding symbolic structure inside LLM representations! Read it if you want to learn how the vectors inside LLMs are like pointillist paintings! news.yale.edu/2026/10/08/h...

It seems that the majority of maths, that is to derive new definitions and theorems, find generalizing structures etc, remains largely human-driven. Trying to prove (long-standing) conjectures might become less meaningful with AI. Was that ever a major part of what mathematicians worked on?

I imagine the step of finding proofs/disproofs of your newly created conjectures used to take up significant time, so that part will be much faster now. Feels overall like a win for the field?

To be clear, I only have a surface-level understanding of the field. So I might be falsely projecting from what happened to ML research and computer science. Any actual mathematicians around to provide better insights?

youtu.beDon't Go Quiet On Me (by Opus 5.5)YouTube video by Neel Nanda

. @neelnanda.bsky.social and Opus 5.5 made a song about MechInterp. It might not be a musical or lyrical masterpiece, but then again which actual artist is going to write a song about this? youtu.be/xhTMRykVb8I?...

arxiv.orgReasoning with Neural Cellular AutomataModern AI architectures used to solve visual reasoning tasks typically rely heavily on global connectivity and synchronization. As biological systems demonstrate, though, sophisticated computation can...

I always like these kinds of alternative neural architectures. New banger paper from folks at Google et al. about reasoning with neural cellular automata. arxiv.org/abs/2609.361...

Daniel Mewes reposted

Imbue

@imbue-ai.bsky.social

What should the future of personal computing look like? We built it 😈 Introducing Imbue Studio: youtu.be/tbdON-tn5gA

youtube.comImbue Studio: the future of personal computingIntroducing Imbue Studio: a new kind of computing environment.Mak...

We're launching Imbue Studio, a personal AI computer for you to use and customize. I haven't had this much fun with using computers in a long time! You can modify any aspect just by asking. Or have it built custom automations and apps. No coding required. Runs locally too. imbue.com/product/studio

The use cases for Sonnet 5.5 still seem pretty niche. It will be cheaper than Opus 5.5 in tasks that require little reasoning and have large inputs, due to lower input token cost. But on all the agentic benchmarks it seems to be at best equal to Opus 5.5 in performance/$?

anthropic.comIntroducing Claude Sonnet 5.5Claude Sonnet 5.5 is a clear upgrade over Claude Sonnet 5, runs 30%+ faster, and costs up to 30% less for most work.

I imagine it produces tokens a little faster than Opus 5.5. So maybe it's faster on some tasks? That could be a use case. Similar to cost, this advantage might disappear when the task requires extensive reasoning or tool calls, where it will need to use more.

Graphs from www.anthropic.com/claude-sonne...

Honest question: What do people mean by "big model feel"? What do I look for in a model's answer to know whether the model is big or not? Does it need specific prompts?

Personally, I haven't found any quality that seems to correlate primarily with model size in my use of LLMs. The type of post-training seems to dominate model characteristics by such a big margin.

If Japan manages to make an economic/technological comeback, I'd honestly not be surprised to see @sakanaai.bsky.social play a big part in making that happen.

sakana.aiSakana AIThe Next Frontier: Welcoming AI Pioneer Jürgen Schmidhuber to Sakana AI

From sakana.ai/schmidhuber/: "This technology is the key to revitalizing Japan’s Monozukuri legacy. [...] these models will serve as the cognitive engines for the next generation of industrial manufacturing [...]"

IMO Sakana has been one of the few labs who have hit a sweet spot between exploring bold off-mainstream ideas (e.g. continuous thought networks, local learning rules), being early in the RSI game (Darwin-Godel machines, ShinkaEvolve), and simultaneously shipping real AI products (Fugu).

arxiv.orgWhy Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention MechanismsDespite impressive visual quality, state-of-the-art video diffusion models often generate content that violates real-world physical laws. While existing solutions rely on external priors or specialize...

Very interesting interpretability work on *video* models by Yueyan Li et al. arxiv.org/abs/2609.23658 "[...] Rotary Position Embedding (RoPE) induces excessive spatial attention decay. This causes early candidate regions to prematurely lock into physically implausible positions."

IMO the biggest improvement in Opus 5.5 is that it speaks in normal language again!

Thanks to @jacksonkernion.bsky.social and team!

Interesting work by J. Seeley and J. Gould at @sakanaai.bsky.social : a local learning rule that has similar performance to backprop in training deep networks. Graph shows comparison to plain predictive coding (PC), which doesn't scale to deep networks. pub.sakana.ai/pc-alm/

The key is to add this Lagrangian per-layer target to the loss function. The Lagrangian multipliers (lambda_i) are vectors that get updated gradually in each inference (forward) step, thereby avoiding the need for a separate backwards pass across layers.

They have some nice animations in the linked post that show how this converges to the same gradient direction as BP after a few inference passes.

AI alignment has been solved folks! "Strong and Smart President is All You Need". Paper coming soon.

13 Sept

What would it look like to train an LLM that is specifically good at dealing with counterfactuals and hypotheticals? Some kind of AI that understands deeply the causal chains behind a given fact? I think this might be the key to AI that's good at novel discovery.

12 Sept

AI capabilities have improved so dramatically over the past 4 years that it's easy to miss that hallucinations are still nearly as much of a problem as they were back then. This is actually holding back LLMs in auto-research: They lack knowledge about *why* they believe a given fact to be true.

12 Sept

Hallucinations have been *partially* solved by providing tools (web search and file reads). But that only works for knowledge that is already available and clearly expressed in the corpus.

11 Sept

Lots of people on X talking about how they "trained the fruit fly brain to do X", not even realizing that back propagation and dotprod+non-linearity activations are not actually how real brains work.

9 Sept

Very interesting work about the shape of CoT reasoning: "[...] reasoning models are dynamical systems with fractal basins, with fractality increasing with task difficulty across diverse tasks." by J. Lai et al: arxiv.org/pdf/2609.04963 I really like this way of looking at LLM traces!

9 Sept

Sharing @kennethstanley.bsky.social 's post on why open-endedness is still very much needed, despite the current pace of new discoveries coming out of existing objective-driven AI systems.

amongai.comLLMs are Crystallized-ConsciousAre LLMs conscious? Today’s LLMs are missing some capabilities that we typically associate with consciousness, such as persistent memory1, continuous updating, or an ability to interact with …

In my new blog post, I argue that LLMs exhibit a particular type of consciousness. Does the idea of crystallized consciousness make sense? Would you consider it interesting, or does it seem completely obvious? amongai.com/2026/09/06/l...

It's actually deeply surprising to me that no lab has been able to build a sustained lead pre-RSI from their internal research. We have 5+ labs continuously releasing roughly capability-equivalent models within no more than 6 months of each other.

Sam Harsimony@harsimony.bsky.social

Pre-registering a prediction that RSI will not produce a durable competitive advantage for OAI/Anthropic. Existing training is already pretty good. RSI will produce a one-time, significant bump that everyone will copy. Then we're back to doing RL on high-quality data. I could be wrong of course!

To be fair, most of the big labs seem to have stopped doing foundational research to explore alternative approaches beyond incremental improvements. With the exception of DeepMind and *maybe* Meta. Small startups (AMI etc.) are still doing it, but might lack resources to pull ahead on their own.

Sam Harsimony@harsimony.bsky.social

And another one: Maybe the difference is that AI labs are so well funded. Thus, once one lab demonstrates that approach X works, all the other ones will pour ~unlimited resources into copying it.

A lot of these factors match past high-profile technology races such as the moon race. I think it's fair to say that USA and UDSSR stayed fairly close throughout the entire process in terms of capabilities. US came out a little ahead, but I suspect their lead couldn't have been more than 1-2 years ?

Sam Harsimony@harsimony.bsky.social

People moving between labs and taking ideas with them is a big factor. But I feel like this is much less prevalent in other tech. Like people probably go from SpaceX to Boeing, and yet Boeing hasn't put out a reusable rocket yet. Are those other areas just slower?

Another interpretation: Most meaningful AI improvements actually still come from academia and are already published. Labs just need to figure out which ones to apply and how to scale them up, so nobody has any secret sauce.