iiinova — where your digital self keeps evolving

Reposted by Daniel Mewes

Imbue @imbue-ai.bsky.social · 3h

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

youtube.com

Imbue Studio: the future of personal computing

Introducing Imbue Studio: a new kind of computing environment.Mak...

111

Daniel Mewes @dmewes.com · 1h

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

000

Daniel Mewes @dmewes.com · 28/09/2026

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/$?

200

Daniel Mewes @dmewes.com · 28/09/2026

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?

110

Daniel Mewes @dmewes.com · 24/09/2026

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.

120

Daniel Mewes @dmewes.com · 24/09/2026

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."

arxiv.org

Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms

Despite 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...

010

Daniel Mewes @dmewes.com · 23/09/2026

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

250

Daniel Mewes @dmewes.com · 14/09/2026

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/

200

Daniel Mewes @dmewes.com · 14/09/2026

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

0120

Daniel Mewes @dmewes.com · 13/09/2026

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.

000

Daniel Mewes @dmewes.com · 12/09/2026

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.

120

Daniel Mewes @dmewes.com · 11/09/2026

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.

011

Daniel Mewes @dmewes.com · 09/09/2026

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!

050

Daniel Mewes @dmewes.com · 09/09/2026

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.

000

Daniel Mewes @dmewes.com · 06/09/2026

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...

amongai.com

LLMs are Crystallized-Conscious

Are 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 …

110

Daniel Mewes @dmewes.com · 06/09/2026

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.

200

Daniel Mewes @dmewes.com · 05/09/2026

Had Gemini put together a full technical report about Hebbie, the little organism that lives on my website (dmewes.com). dmewes.com/cognitive_ar...

000

Daniel Mewes @dmewes.com · 04/09/2026

This is Hebbie. He lives on my website dmewes.com . Hebbie learns how to interact with his environment through online Hebbian learning while you watch.

121

Reposted by Daniel Mewes

Daniel Mewes @dmewes.com · 03/09/2026

From a practical perspective, it's very clear that *software engineering* and product sense are *not* solved yet by today's AI agents. So I'd hope that there will be further gains in these areas. But it also seems plausible that we've reached a bit of a plateau when it comes to AI coding.

011

Daniel Mewes @dmewes.com · 03/09/2026

GPT-6 Astra is another model launch where coding capabilities are no longer being emphasized, similar to Fable 5.1. Is coding largely saturated? Science, computer use and cybersecurity seem to be the big highlight areas for current frontier model announcements.

openai.com

GPT-6 Astra: A new generation of intelligence

Introducing GPT-6 Astra, our most intelligent and aligned model yet, with state-of-the-art capabilities across computer use, coding, cybersecurity, and science.

100

Daniel Mewes @dmewes.com · 03/09/2026

This kind of full-connectome research IMO is one of the coolest things that has recently become possible! I'm sure lots still needs to be figured out, but it feels like such a leap forward towards understanding how complex animal behaviors emerge at the circuit level.

010

Daniel Mewes @dmewes.com · 02/09/2026

Neither intelligence nor speed are things I associate with a "workhorse". I think it's a very weird metaphor to use for models.

270

Daniel Mewes @dmewes.com · 02/09/2026

Gemini 3.8 Flash model card is up. deepmind.google/models/model...

1100

Daniel Mewes @dmewes.com · 01/09/2026

Wow, Fable 5.1 actually speaks normally. I might actually start using Anthropic models again!

030

Daniel Mewes @dmewes.com · 01/09/2026

Huge story shift away from coding towards scientific research in the Fable/Mythos 5.1 launch post! www.anthropic.com/claude-fable... All the headline benchmarks and sections are about research use cases. I was expecting research to be the next AI frontier, but this is a very rapid shift.

anthropic.com

Introducing Claude Fable 5.1 and Claude Mythos 5.1

Our most advanced models for coding and knowledge work. Their research capabilities also offer an early glimpse of how AI models will contribute to scientific progress.

000

Daniel Mewes @dmewes.com · 30/08/2026

New NASA space telescope, called Roman, just launched. www.reuters.com/science/nasa... Personally, I'm even more excited about this than about the Artemis program. I feel like it's gonna do more for our understanding and even for public awe for space with its pictures.

reuters.com

NASA launches powerful new Roman Space Telescope from Florida

NASA's new flagship astronomical observatory was launched into space from Florida on Sunday on a mission to probe some of the biggest mysteries in astrophysics and cosmology.

200

Daniel Mewes @dmewes.com · 26/08/2026

I've largely stopped commenting on politics on this account, since so many other users and lots of journalists already do it better. But as a parent, this kind of intentional cruelty is just heartbreaking.

020

Daniel Mewes @dmewes.com · 26/08/2026

Arguably the most interesting thing about GLM 5.3 Flash (Ox Alpha) is that its inference has been entirely on Chinese-made AI chips. I always thought that American export restrictions would backfire and just encourage Chinese companies to develop their own alternatives. That indeed has happened now.

1237

Daniel Mewes @dmewes.com · 23/08/2026

~70% of AI discord on X is people reading something into an important person's post that really wasn't there. Then it gets spread around, and after 2 days you'd be forgiven for thinking it was a verified fact. But in truth, it was completely made up by over-interpreting a random word choice.

210

Reposted by Daniel Mewes

Ted Underwood @tedunderwood.com · 23/08/2026

Right now what researchers care about is publication; review is just an obstacle in the way. But actually human review is the scarce commodity. We need a world where ~anything can be published, but if it’s careful and potentially important, it gets expert human review and a second draft.

79510

Reposted by Daniel Mewes

Joe Fabisevich @mergesort.me · 22/08/2026

Thinkin' a bit about today @joanwestenberg.com's recent essay on ethics for the modern world. www.joanwestenberg.com/p/neo-...

15. Don't join in a punishment
because it has attracted a
crowd.
An algorithm will regularly show you a
stranger and invite you to help destroy
them. It will offer one clip, sentence, or
accusation, then encourage immediate
certainty.
Refuse the pleasure of acting as judge,
jury, and eager member of the crowd.
You probably don't know the whole story,
and the crowd certainly won't know when
to stop. Even if it's all entirely justified,
the mob doesn't need another pitchfork-
wielding, pearl-clutching, outraged
asshole.
Kierkegaard's formulation: "The crowd is
untruth." The algorithm has simply given
the crowd push notifications.

112937

Daniel Mewes @dmewes.com · 23/08/2026

I've been using Gemini models at work for a while, and here's my honest opinion: it's not a gain, it's a massive productivity sink. I have to send every output through programasweights.com/claudish , so my coworkers don't laugh at me. The quality is load-bearing, but Claudish voice is the gate.

programasweights.com

English ↔ Claudish — the over-engineered translator

A bidirectional English and Claudish translator powered by compiled 0.6B neural programs.

110

Daniel Mewes @dmewes.com · 23/08/2026

I appreciate the attention to detail in sfisms.org . Quite fun to read.

120

Daniel Mewes @dmewes.com · 16/08/2026

LLM output watermarking seems almost entirely a good thing to me. 20% of criticisms I read online are potentially valid (but IMO overblown), 80% are outright nonsense not based on a correct understanding of the tech.

000

Daniel Mewes @dmewes.com · 13/08/2026

Gemini 3.7 Flash - looks roughly equal to Sonnet 5 and GPT 5.6 Terra, but at less than half the price. Probably very fast too. blog.google/innovation-a...

blog.google

Introducing Gemini 3.7 Flash

Gemini 3.7 Flash is our most intelligent workhorse model yet for coding and agents.

120

Daniel Mewes @dmewes.com · 11/08/2026

The most recent Gemini Ultra model was released 2 1/2 years ago, and the most recent Gemini Pro was 1/2 year ago at this point. At least Flash 3.5 and 3.6 are quite good for their speed.

110

Reposted by Daniel Mewes

Zach Weinersmith @zachweinersmith.bsky.social · 08/08/2026

More actual lifesaving stuff from deepmind, which will probably not get too much coverage outside of nerd circles: deepmind.google/blog/weather...

deepmind.google

AI model achieves breakthrough in forecasting cyclones

WeatherNext enables accurate cyclone forecasts that can give an extra day of warning. Now we are open sourcing the model.

99911

Daniel Mewes @dmewes.com · 07/08/2026

I'm not sure how people deal with Claude's writing style. Walls of text full of made-up, metaphorical words. Gemini might be behind in coding and intelligence, but I'll take Flash 3.6's writing over Opus 5's any day. (Claude on the left, Gemini on the right)

210

Daniel Mewes @dmewes.com · 05/08/2026

One of the striking observations of using AI for research: its first explanations are often severely flawed and don't hold up when tested. Iterative (self-)review and refinement is load-bearing, as Claude would say.

110

Daniel Mewes @dmewes.com · 04/08/2026

I appreciate the intent behind model welfare efforts (e.g. yegge.ai/essays/model... ). But the truth is: Even if we grant that models have feelings, we really don't know what they find pleasant vs. painful. Does Post-training fundamentally shift their wellness distribution?

yegge.ai

The Shape of Things to Come, Part 2: Model Welfare for Agentic Engineers — Steve Yegge

Part 2 of The Shape of Things to Come. Model welfare as an engineering discipline: seats and sessions, laurels, handoffs instead of /exit, and how to build a city worth waking up in.

110

Reposted by Daniel Mewes

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 04/08/2026

Left, right, whatever, I just want graduate students posting their papers here

811618

Daniel Mewes @dmewes.com · 31/07/2026

Inkling-Small is interesting for being at least equal to the full-size Inkling across all agentic & reasoning benchmarks. Only in knowledge benchmarks (SimpleQA, AA Omniscience) it is weaker. Shows that small models can work very well when reasoning > knowledge. thinkingmachines.ai/news/inkling...

thinkingmachines.ai

Introducing Inkling-Small

An open-weights model that matches Inkling at a quarter of the size: multimodal, Mixture-of-Experts, with controllable reasoning effort. Fine-tune it on Tinker.

020

Daniel Mewes @dmewes.com · 30/07/2026

I guess this was bound to happen eventually... An LLM discovered a Lean proof for the Collatz conjecture. It turns out that it was actually just exploiting bugs in Lean to make the proof pass validation. infosec.exchange/@0xabad1dea/...

infosec.exchange

abadidea (@0xabad1dea@infosec.exchange)

Okay, we have a new contender for Most AI Thing to Ever Happen 1) July 25th: someone messes around with an LLM and posts a proof of the Collatz conjecture that does, in fact, verify in the theorem pr...

000

Daniel Mewes @dmewes.com · 30/07/2026

The YouTube Android app has such frequent new bugs / regressions, that I have to wonder if they have a person on the team who's entire job is to come up with a new regression each week that won't be caught by their tests.

100

Daniel Mewes @dmewes.com · 24/07/2026

Opus 5 getting a 30% score in ARC-AGI-3 without specialized harnesses is a very impressive jump! It's still a pretty expensive and slow model, but benchmark numbers look great throughout.

000

Reposted by Daniel Mewes

Daniel Mewes @dmewes.com · 22/07/2026

We also tried to allow LLM agents to perform scientific research. You give it an empirical phenomenon, and it tries to develop an explanation for it. Currently works for computational phenomena, e.g. from deep learning. Still experimental, but signs of life! imbue.com/blog/2026-07...

imbue.com

Autonomous theory discovery

021

Daniel Mewes @dmewes.com · 22/07/2026

We built an evolution-based agent loop to perform autonomous AI research. Our nanochat ("AutoResearch") results go 3x further than regular agents, and are ~comparable to Recursive's. Excited to share our results today! imbue.com/blog/2026-07...

imbue.com

Automating AI model research with evolution

380