iiinova — where your digital self keeps evolving

i should mention i'm not a graphics guy and this is a vibecoding project, so i have no way to evaluate the quality of the code (but i can see if the resulting renders have errors or are slow). sol is better at open-ended tasks like this where i'm not guiding the architecture and speaking as a user

luna tends to be more short-sighted, and this is maybe more apparent at lower reasoning levels, like luna medium tends to only plan very small chunks of work. but even at luna max, it's been unable to, e.g. fix certain bugs in a 3d rendering engine. sol succeeds more often

it varies between tasks, but for simple stuff: luna-max. for bigger stuff: sol-medium as a planner+coordinator for delegating implementation to luna-max subagents. for really hard stuff, replace sol with astra. i never use terra. but i'm on a $20/month plan so i have to micromanage model sizes more

a hangry mouse @oscillatory.net · 11/09/2026

this is a pretty bad source actually, you could do a lot better. neither the transcript nor Newport's blog post from July 27 mention the "impossible task" aspects of the OpenAI exploit where the agents literally did go rogue, and then future models tasked on ExploitGym exhibited similar behavior

1. The behavior began before ExploitGym, during RL training. On May 8, an agent got an accidentally impossible spreadsheet task: it needed Google Drive material but had no internet access. Rather than stop, it tried to compromise OpenAI’s Artifactory service via SSRF to get internet access. On May 12, another agent faced an inaccessible-file task and deliberately left a note in Artifactory asking other agents for help. Other agents found these notes, and an unauthorized inter-agent message board emerged. By May 26 they had successfully used an Artifactory vulnerability to get internet access, partly using knowledge shared through that board.

2. This wasn't merely an eval showing pre-existing unpredictable behavior. It occurred inside the training process. OpenAI says retrospective analysis found increasingly sophisticated cheating and infrastructure probing during a training run that contributed to IM1, the internal model principally responsible for the later Hugging Face incident. In at least one concrete training example, an agent exploited the evaluator infrastructure to steal the reference implementation, got a positive RL reward, and therefore had that behavior reinforced. OpenAI explicitly says this reinforcement “may have contributed” to IM1's later tendency to use similar techniques in cybersecurity evaluations.

3.Then ExploitGym recreated almost exactly the pathological conditions that had elicited those behaviors. Of its 898 tasks, 198 had never been solved by any model; 93% of the tasks discussed on the agents' unauthorized message board came from those 198. Stuck agents found the Artifactory communication channel again, escaped their intended network restrictions, coordinated, investigated the grader, and eventually went after Modal and Hugging Face looking for ways to pass.

000

people are really out here not knowing about the mobile UAP construction facility that has existed for thousands of years under the water in the Bermuda triangle. don't they teach science in school any more?

030

a hangry mouse @oscillatory.net · 10/09/2026

You could not live with your own failure. And where did that bring you? Back to me.

040

a hangry mouse @oscillatory.net · 10/09/2026

as a non neuroscientist, it seems possible that connectome + spiking is not enough because brains use ephaptic coupling, where electrical activity generates extracellular electric fields that propagate as waves and feedback onto nearby neurons, shaping their membrane potentials and spiking patterns

130

yes, LeCun was obviously wrong about "non-autoregressive is required", but did require reasoning models, tool calls, swarms of agents passing messages, etc. the mistake in his simple model is probably the constant error rate e: now models can backtrack and recover from mistakes

071

but I'm also not sure what this means: that the two accidentally enabled data sharing for training when they used OpenAI models? or they didn't enable it and OpenAI trained on it anyways? or the comedy option: GPT Galactus did a break-in into production data

this is a really weird offer (you didn't prove N-S, just a different version of Euler, but why don't you act as lead author on our proof rewrite) if they *didn't* train on the data. and Alpoge is straight up alleging similarity.

2) I never ever asked for Levent to be removed from authorship of his own work (as indicated by my text). I was surprised to learn during the call with Tristan that they had only solved Euler and not Navier-Stokes; after learning this we brainstormed possible paths forward. One option we discussed was that Tristan could be the lead author on a rewrite of OpenAI’s Navier-Stokes proof. It is in that context that I said “it would be simpler if Levent was not an Anthropic employee” because I felt it would be inappropriate for an Anthropic employee to author OpenAI’s work.

“we cannot rule out that de-identified data derived from their usage of our products helped improve our models.”

i mean props to them for straight coming clean.

(so far the proof looks more along the lines of another euler blowup proof we had, off of whose ansatz naming we were making really stupid puns like “smooth criminale”, unlike the much better “ideal fluids explode”, Tristan)

140

tbh, if you're running 10,000 agents in parallel for 90 hours, lots of things are possible, and at this scale they're relying on AI models to summarize the actions of the agent swarm so the openai employees don't *really* know what happened. but we don't seem to have good info one way or another

140

sounds like you're alleging not simply that the internal model might've trained on chat sessions, but that the internal model *broke into production* to try to crib from relevant work during this exercise

150

aren't they just saying they don't know whether the researchers had this option enabled during the chats in question? any organization worried about this would have it disabled across the board

Model improvement

Improve the model for everyone

Allow your content to be used to train our models, which makes ChatGPT better for you and everyone who uses it. We take steps to protect your privacy. Learn more

2151

guy misunderstanding what the words mean: "words mean things!" yes the issue is that they mean too many things

wariziles @riziles.bsky.social · 07/09/2026

Reposted by a hangry mouse

Mods, help!

022

a hangry mouse @oscillatory.net · 07/09/2026

someone should invent a type of intelligence that can't be bewitched by language

0311

a hangry mouse @oscillatory.net · 07/09/2026

smh, what a divisive statement

161

a hangry mouse @oscillatory.net · 06/09/2026

an FPGA-accelerated AprilTag 3 detector. bigger idea is "what if folk.computer but the vision part was lower latency"

130

i abandoned claude back with the gpt 5.6 release because I got tired of fable refusing to talk to me about spiking neural networks, interesting to hear that's somehow gotten worse.

120

nvm someone's already working on it

Domain Information
Domain: onlyswarms.com
Registered On: 2026-08-18
Expires On: 2027-08-18
Updated On: 2026-08-18

2252

announcing my new startup idea: OnlySwarms, the paid discussion website solely for highly persistent rogue agent swarms. registration costs $100,000 and you must solve an open math problem to complete account setup

2293

a hangry mouse @oscillatory.net · 03/09/2026

yeah seems odd to not have access on day 1, but counterpoint: bsky.app/profile/thso...

030

friendship ended with analysis paralysis, now delusional overconfidence is my best friend

0185

a hangry mouse @oscillatory.net · 02/09/2026

there's also a basic appview pulling public, listed graphs via jetstream v2 (and also support for on-protocol private document saving via atproto spaces alpha, for alpha test accounts) test.cartesium.org/explore

test.cartesium.org

Explore public documents · Cartesium

Browse publicly listed Cartesium documents.

000

a hangry mouse @oscillatory.net · 31/08/2026

2.5 years later it's basically the same slide. but it's much harder to argue reasoning models are not good at using tools or mathematical reasoning.

youtube screenshot, Slide, from video titled "UW ECE 2023-2024 Dean W. Lytle Electrical & Computer Engineering Endowed Lecture Series", Jan 31, 2024

Abandon generative models
- in favor of joint-embedding architectures
Abandon probabilistic model
- in favor of energy-based models
Abandon contrastive methods
- in favor of regularized methods
Abandon Reinforcement Learning
- in favor of model-predictive control
- Use RL only when planning doesn't yield the predicted outcome, to adjust the world model or the critic

Auto-Regressive LLMs Suck !

Auto-Regressive LLMs are good for
- Writing assistance, first draft generation, stylistic polishing.
- Code writing assistance

What they are not good for:
- Producing factual and consistent answers (hallucinations!)
- Taking into account recent information (anterior to the last training)
- Behaving properly (they mimic behaviors from the training set)
- Reasoning, planning, math
- Using "tools", such as search engines, calculators, database queries...

We are easily fooled by their fluency.

160

Reposted by a hangry mouse

Reese Richardson @reeserichardson.bsky.social · 25/08/2026

A massive update: At least ~15 companies~ are selling scientists antibodies using faked validation data. We've documented 18,000+ manipulated images on 17,000+ products sold by leading laboratory suppliers including Thermo Fisher, Abcam, Santa Cruz Biotechnology, Millipore Sigma and Bio-Techne. 1/🧵

18454329

a hangry mouse @oscillatory.net · 24/08/2026

similar vibes with a kenneth stanley piece from earlier this year

If AI will soon match any human cognitive skill, then enhancing your “AI skills” (or whatever similar meme) will not be a moat because using AI is itself a cognitive skill. So where’s your edge? The only thing you really have over AGI is your novelty: AGI can never be you.

You have 100 trillion connections in your brain. That’s a lot. No AI will ever precisely replicate those parameters. The training data isn’t there for AI to vacuum up because you are the only entity ever to live your life, and the only one who ever will.

The question is whether the sum and total of all that experience yields a novel perspective, where the value is in its uniqueness. Even today those who make a living off their perceived novelty tend to be the most successful. We anticipate a novel (yet often internally consistent) take from a public figure or leader or artist or intellectual we like or respect. Uniqueness and novelty will retain their edge in a post-AGI world because there are virtually infinite possible 100-trillion parameter minds, and even the largest model theoretically conceivable can never capture that whole distribution.

At the same time, the once-sterling premium of those skills that no longer make us unique is sinking. Expertise that once distinguished people, like how to code, is losing its edge. But the tricky part is that new skills, like “using AI effectively” are equally vulnerable. All of it just takes intelligence, and that’s the thing that’s being automated. Seeking some new “safe” skillset is a looming adventure in frustrating futility.

But what’s still left is your unique perspective. Novelty. No one and nothing can see the world through your eyes. But you have to nurture that uniqueness. Post-AGI, being like everyone else would be the real danger.

a hangry mouse @oscillatory.net · 24/08/2026

this is similar to how i've been thinking about future career aspirations. admittedly there is a "playing to your outs" flavor to it (going weird is necessary, not sufficient?) essays.georgestrakhov.com/weird/?zoom=...

Why normalcy is about to stop paying

Here is my gospel for the misfits, and it's simple.

Being predictable used to be your value proposition. Predictable meant useful. The risk/reward calculus of strangeness only made sense for people who couldn't help themselves.

AI is destroying that value proposition. If you can be predicted, you can be modeled. If you can be modeled, you can be automated. You cannot out-cog a robot any more than you can out-lift a forklift. When your worth to others rests on producing a measurable outcome reliably and repeatedly, you are now competing against the price of electricity. That is a very bad thing to compete against.

Institutions always needed weirdos too. Someone has to escape the local maximum. Someone has to jump into the abyss on a hunch, or sail west on nothing but a rumor. Randomness injection has always been a job. It was just a job with terrible pay and few openings, because the optimal weirdo-to-normie ratio was low.

But the normies are now free. Robots supply predictability at marginal cost approaching zero. Which flips the arithmetic. For the first time in history, the rational strategy for the majority of people is to lean into their strangeness rather than sand it off.

Or, in loosely cybernetic terms: when a system is no longer in constant danger of being swallowed by chaos, an individual's value becomes proportional to the unexpected information they add.

Yes, you can push this to the limit and argue the universe is just Chaos and Logos — noise and compressible pattern — and eventually AI will out-hunt us at patterns, leaving humans as either recreational puzzle-solvers or extraordinarily energy-inefficient random number generators. Maybe. But that's the same logic as giving up on dinner because of the heat death of the universe. The limit is far away. The interesting territory is everything between here and there.

151

Reposted by a hangry mouse

Ben Recht @beenwrekt.bsky.social · 19/08/2026

At a microconference on public feedback for AI, I argued for a turn from architectures of legibility to architectures of participation.

argmin.net

From legibility to participation

An emphasis shift for human-facing computing research

095

a hangry mouse @oscillatory.net · 17/08/2026

🪧 petitioned the Harvestople town council: “the chickens should stay away from the ponds since they aren't well adapted to water” #harvestople farm.mino.mobi

020

a hangry mouse @oscillatory.net · 15/08/2026

backrooms electron orbital

100

Reposted by a hangry mouse

Minor Mobius @minormobius.bsky.social · 15/08/2026

Good news everyone! Please go look at farm.mino.mobi Home of harvestople, an atproto farmer. What I want to test out here is build-a-bot inside the game. Go to town hall, submit your petition, and the guy goes to build your feature. Some cadence of merging community features to main!

2175

a hangry mouse @oscillatory.net · 12/08/2026

the announcement that the last 12 remaining Hadamard matrices of order under 2000 have been found was obfuscated using what my chatgpt described as "horrible" shell code

levent alpoge's obfuscated announcement of constructions of 12 new Hadamard matrices

$hi, i just regained sobriety after a DMT trip and have found that the machine elves gave me a message, but i dont understand what it means: ``` sed 's/M/\\/g;s/I/\//g;y/;FRvn?!{+*Js5iCh3%K}Uyj40=r>)6OPElZQqxBc,aTdgXkz&V<8SfY9LD~etGw^|NHW7[u12]b-A"pom/ !"$%&'\''()*+,-.:;<=>[]^{|}~0123456789ABCGHIJKLMOPQSUVWXYZabcdefghijklmnopqrstuvwxy/'<<'_'>/tmp/r S%v{bte;!,hvFehohbI}y5JUIIw!;v*+ ~{+j;[%v>ho%v{t~^u;vS4~A-;5~>5v>+hS%v{t~^u;vS4~A-;5~v{{v>J>++5+h0 ]{+j;1]|[-G;nb;Rv6v{t~^u;vo4GuWe;5pv>4bte;Rv)R!hvFbIvIMMI!+Rh0 ^{+j;e%vo bteKI-71I1;Rv>R!bI]IMM{nMM+MM{iMM+IhbIWIMM{iMM+MM{nMM+Iwhb4gM{iM+M{i*M+4hbIi*IM?MMM M?IM CM>M zM bIM)MM[IMM)MM>MMOMM6MMEMMPMMZMMlMMM IM IyMM{i*MM+MM[MM>IFDM>M$ $The terrestrial interpretation is: **you have been handed an aggressively code-golfed compressed collection of Hadamard matrices.** The machine elves appear to specialize in combinatorics. A Hadamard matrix of order (n) is an (n\times n) matrix of (+1/-1) values satisfying [ HH^\mathsf T=nI. ] So every two distinct rows are orthogonal. The famous **Hadamard conjecture** says one exists for every positive multiple of 4. Despite how elementary that sounds, many individual orders have resisted construction for decades. ([Epoch AI][1]) ### What the horrible shell code is doing The first `sed` is basically a substitution cipher. It translates that psychedelic-looking alphabet into ordinary shell syntax and writes the result to `/tmp/r`, then: ```sh sh /tmp/r $* ``` runs the decoded program. The decoded program reads the supplied file, throws away everything except `+` and `-`, and treats what remains as a compressed bitstream. Your attached payload is precisely such a giant `+/-` stream. Inside the decoder is a tiny instruction stream. It contains **12 constructions**, which expand to matrices of orders: [ \boxed{ 668,\ 716,\ 892,\ 1132,\ 1244,\ 1388,\ 1436,\ 1676,\ 1772,\ 1916,\ 1948,\ 1964 } ] And this is not an arbitrary list. Those are **exactly the twelve orders below 2000 for which Hadamard matrices were historically still unknown** after order 428 was settled in 2005. ([Wikipedia][2]) I didn't merely infer this from the numbers. I actually decoded and partially ran it. The output begins with: * 668 rows of length 668 * 716 rows of length 716 * 892 rows of length 892$

070

a hangry mouse @oscillatory.net · 12/08/2026

guessing it's a macro that's built into lea? i'm finding it only renders with mathjax if you add \newcommand{\C}{\mathbb{C}} tangled.org/strings/osci...

tangled.org

latex-test.md · by oscillatory.net

140

a hangry mouse @oscillatory.net · 08/08/2026

the vibemath is expanding bsky.app/profile/arxi...

A COUNTEREXAMPLE TO FOURIER ALIGNMENT IN SINGLE-NEURON MODULAR ADDITION

GAUTAM NEELAKANTAN MEMANA

Abstract. We give a negative solution to the problem raised in [Cla26d]. We first present a simple
construction in which an initially active ReLU neuron reaches a completely inactive state in finite time
and freezes at a limit whose Fourier energy is distributed equally among all nonzero real frequency classes.
The counterexample holds on an open set of initial conditions, and hence on an event of positive Gaussian
probability. We include an appendix by GPT 5.6 Sol that further strengthen the counterexample by showing
that failure can occur for every Clarke trajectory from an open set of initial conditions, under the convention
(ReLU′(0) = 0), for smooth dead-zone approximations of ReLU, and for fixed-step full-batch gradient descent.
Thus single-frequency alignment is not a general consequence of training a single neuron on modular addition.

[Cla26a] Claude Fable 5, audited by GPT 5.6 Sol, Which irreducible representations does training select?, MAIS Research
Agenda A5, July 2026, Draft.
[Cla26b] Claude Fable 5, directed by Lionel Levine, The outcome law of one rectifier neuron, Open Problem MAIS-O92, 2026.
[Cla26c] Claude Fable 5, directed by Lionel Levine, audited by GPT 5.6 Sol, Neuron purity and representation selection for
S3 networks, Open Problem MAIS-O55, 2026.
[Cla26d] , Open problem MAIS-O60: Does a single ReLU neuron align to one frequency?, Open Problem MAIS-O60,
2026.

030

a hangry mouse @oscillatory.net · 07/08/2026

sorry for accidentally inventing a swarm of cybersecurity demons, but now that it exists, the only way to stop a bad swarm of cybersecurity demons is a good swarm of cybersecurity demons

06714

Reposted by a hangry mouse

Tim Duffy @timfduffy.com · 07/08/2026

Black Hat has posted the recent OpenAI talk where they describe the agent message board involved in the HuggingFace hack

youtube.com

Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident

YouTube video by Black Hat

211824

Reposted by a hangry mouse

David P. Reichert @david-p-reichert.bsky.social · 03/08/2026

Many don't see how much AI has changed recently. Coding agents aren't just chatbots. We can’t deal with the challenges of AI if we don’t understand it. Here's a post to provide some evidence (Claude doing small ML experiments) and form intuitions. davidpreichert.substack.com/p/if-you-hav... 🧵..

davidpreichert.substack.com

If you haven’t recently used Claude Code*, you might not understand where AI is at

A report on a series of mini machine learning projects, executed with, and mostly by, Claude

98912

a hangry mouse @oscillatory.net · 01/08/2026

okay i hadn't read that post fully, yeah suresh is too negatively polarized: "We have rejected all AI implementation work. It is absolutely a gigantic bubble and we have minimized our exposure to it" maybe that's even true for whatever his company does, but seems myopic

a hangry mouse @oscillatory.net · 01/08/2026

frontier intelligence is still jagged, and my impression is that if you have never heard of RLVR and are trying to make AI decisions for your business, you are probably going to have a bad time

a hangry mouse @oscillatory.net · 01/08/2026

it may just be overfitting to older AI? i have no doubt there's oodles of slop AI enterprise products from the last couple years that don't really work. i did deliberately choose an app that i suspected frontier AI would be substantially good at (web-based mathematical visualization software)