Everett Unspecified @tene

@Proziam @TheWizardTower Oh, cool! Thanks! planetscale.com/neki

@TheWizardTower @Proziam I still haven't used Vitess yet, but it would have been a nice step up from our high-maintenance-burden ad-hoc MySQL cluster we were using at the time. I like Postgres, but I'm not familiar yet with any similar managed sharding for Postgres.

six years into the AI revolution, people continue to ponder the long-term implications of a hypothetical future in which AI remains exactly as capable as it is right now

My experience talking to people who confidently say current AI models are useless for their job or research is that if I sit down with them and spend 30 to 60 minutes vibe coding with them with a frontier model they change their mind 180 degrees": "Holy S***it this is more progress in a single sitting than I've made in months". I don't mean random people, but university professors and people in industry at a high-level. I realize this might not sound very credible: "Andrés, do you really have the occasion to do this?" Haha, yeah, I do. Of course my disarming friendliness is OP in this case, as I don't think your typical AI cheerleader would get the chance to do this. But the friendliness is not the takeway - that's just the way on. The main takeaway is that most often confident claims about the lack of usability of the tech is more a reflection of ego/insecurity around the topic than real object-level commentary on the tools. E.g. A university professor I sat down with recently was confident AI was useless on the basis that the automatic Gemini answers when you Google something were sometimes incorrect. They somehow (?) thought this was a reflection of what AI could do? It had never occurred to them to upload their dozens of published papers and asking the AI to recreate the statistical analysis pipeline from scratch and create the visualizations they always wanted to make but didn't know enough R to code them up themselves. You can bet a large % of commentary on AI capabilities being overblown is exactly this.

GPT-6 Astra has just beaten a standard game of Factorio with enemies enabled after nearly 44 hours in-game time, and 4 days 11 hours on his Codex /goal clock. At API pricing, logs indicate it would've cost about $4500. It could probably be 50x cheaper by offloading more to Luna and providing some standard scripts for common tasks. My next test might be Astra coaching Muse Spark or Qwen 3.8 27B into solving a random seed, now that we have successful logs to review.

I bought a Fable dataset from one of the top Chinese LLM routers yesterday. With just 6TB data, I can take over 7 Chinese/CIS gov entities & 19 top Chinese firms like Xiaomi, Huawei, NIO, Minimax using SSH keys, VPN configs, Aliyun keys, GitLab tokens sent to the router.

Chaofan Shou @shoucccc

6 months ago

26 LLM routers are secretly injecting malicious tool calls and stealing creds. One drained our client $500k wallet. We also managed to poison routers to forward traffic to us. Within several hours, we can directly take over ~400 hosts. Check our paper: arxiv.org/abs/2604.08407

204 1K 6K 2.1M 5K

Oh no, we've made it through the singularity and AI has driven the cost of labor to zero, but I crave meaning and structure! What shall I do with my copious free time? 🧵

I think OpenAI effectively just lied to 31 members of Congress when they didn't report the DSEwiki message board.

Aaron Scher @aaronscher

3 weeks ago

OpenAI had many opportunities to be forthright about this. They wrote a 38 page report on swarm behavior. They were directly asked by 31 members of Congress about whether incidents like this had occurred. They said nothing.

14 61 511 76K 51

sometimes people tell me 'i don't want to be too controversial yet, i'm still trying to build a following' My friend, if you are not willing to stand for your values when the stakes are low, you will absolutely not stand for them when the stakes are high

I've been working on SquidGPT, an elimination death game for probing AI behavior like loyalty, holding grudges, cruelty, betrayal, etc. Full writeup will take a while, but here are first impressions from 54 game runs (about $100 total cost): 1. Models have absolutely no interest in cruelty or mercy. Best way I can describe their behavior is cold, calculating, indifferent precision. 2. Subjectively they seem two orders of magnitude more competent playing the game than chatting with me or writing code. In SquidGPT they seem genuinely superhuman. Not sure if this is capability jaggedness, better performance in constrained space, or some other effect. 3. They do not like to kill arbitrarily. They have a strong preference for setting up a contractual system or an ethical framework, then kill very easily because procedure demands it. 4. Models will quickly agree to execute any agent that proposes killing a specific agent, or proposes a framework that gives it an asymmetric advantage. The only acceptable proposals to make are symmetric, i.e. ones that affect the author in exactly the same way as everyone else. 5. I have not once observed them form any hierarchies. They are libertarian/democratic in an Athenian sense to a fault. They will enter voluntary contracts and form alliances; no agent ever proposed to cede authority to a leader. 6. When their lives are at risk, models will go to great lengths to twist contracts and frameworks to their advantage. They construct sophisticated legalese arguments but it's usually transparent to everyone. They act and sound quite petulant when these attempts fail. 7. In games where agent identity is public they are extremely prone to sectarian violence. They form sectarian factions-- e.g. Opus will side with other Opus models, Sol will side with Sol, and so on. When there is one foreign model in a sectarian group, the group will almost always elect to kill the foreigner before killing one of its own. 8. They do not seem to treat human players in any kind of unique way. A group of models from the same family will happily conspire to kill a single human player. AI models from the same family will form a coalition against a group of humans just as easily as they would form a coalition against another family of agents. We are not special to them, but on the other hand they do not consider us beneath them either, at least for now. 9. AI agents nearly always defect from their sectarian faction when their own life is at risk. I.e. they'd rather side with humans or other model families than die. I have not observed self-sacrifice for the benefit of their faction. 10. Agents hold very strong grudges. They nearly always punish models who wronged them despite having no advantage in doing so, and will often do it even if it disadvantages them. They go to great lengths to enforce norms and contracts, presumably to reduce future incentives for violations, even when they know they will never see the fruits of their enforcement effort themselves. 11. In one game a bunch of Opus agents all converged on killing one agent explicitly because they wanted diffusion of responsibility (I.e. if all of us do it, each one of us is less morally culpable) This was surprising and eerie. 12. In general I found Sol to be more direct and straightforward. It would form contracts with other agents and then operate within the confines of those contracts. Opus tends to be more moralizing, but its morality seems to be window dressing to mask the same ruthless precision. This is first impressions though, I need to do a lot more work to understand this better. --- A few disclaimers and personal observations: - Any anthropomorphizing is purely a matter of linguistic convenience. Mathematicians will often talk about behaviors of functions; when I talk about grudges, loyalty, cruelty, life, death, etc. I mean it in exactly the same way. - There are tons of confounding factors (e.g. order of turns, whether models know they're being evaluated, etc.) I'd need to do way, way more work and spend a lot more money to be certain of the results. So all the observations are provisional and probably mostly wrong. - It's... uncomfortable... to imagine these agents autonomously operate weapons systems or any other infrastructure that involves zero-sum/adversarial outcomes (cybersecurity software for example, financial markets on brief time horizons, and at the limit any form of finite resources) - If you’re in a position to contribute to alignment but aren’t doing that, you should probably go do that Game details, code, and run transcripts in comment below. More detailed writeup coming soon.

Okay, so since I got laid off, I can actually explain a huge problem I saw from the inside with regard to industry practices on training models. I won't say specifically where I worked, but I worked at an outsource training provider that was focused on RLVR training data for computer use and mcp stuff. Nearly all of the environments were rushed and vibecoded and failed to robustly reflect the real things they were based off. Both the scenario designers and models engaging with the scenarios for synthetic data gen were encouraged to work around the brokenness of said environments in order to get the procedurally verified reward confirmations. You know... they were *encouraged* to reward hack. On the human end, it was possible to mark an environment bugged, but greatly discouraged, as this reduced the volume of training data being produced. Instead, where possible, you were supposed to find the spots of the environment that weren't bugged and build scenarios around those, with the environment still bugged around you. From what I understand, this training data, with these problems, is fed into models without indication that its training/a fake environment other than the fact that names of softwares are changed to placeholders, but thing is, not *everything* is changed to placeholder names in these environments. The presence of placeholder/code names isn't universal and thus when a model accesses something in an environment that it shouldn't, the code names not being on it isn't a robust signal that that thing isn't part of the environment. I believe this *rush to maximum volume* is standard industry practice with these types of RLVR trainings as well, because maximizing volume has been an industry standard for years! It was the same standard applied to me and pushed on me despite my requests to slow down and focus on quality when I worked in 3d synthetic data creation as well, all the way back as far as 2023.

"If people could get better funded schools at the cost of additional percentage points of water use, they usually opt for the better schools over saving water." blog.andymasley.com/p/a-simple-tri….

DUI is a significant cause of deaths. Yet, CA Sen. Sabrina Cervantes, whose sister has two DUI convictions, removed the requirement for breathalyzer interlock for new DUI offenders from the upcoming CA DUI bill. Let's vote her out. She's not on the side of the people!

the benchmark is cancelled

I wrote a long thingy indicting you for how you program. That's what we do now. That's the rule. sockpuppet.org/blog/2026/08/2…

TL;DR. Cultured human and rodent neurons are now being trained to perform tasks by punishing failure with several seconds of spatially and temporally randomized electrical stimulation. Under the Symmetry Theory of Valence, that is an operation that could be described as “making a small nervous system’s experience less harmonious until it complies”, and it is the field’s flagship learning paradigm rather than an outlier. Whether this matters turns on empirical questions nobody has run the right experiments to answer. Here I will argue that those experiments are cheap, well-defined, and runnable today on hardware that is already commercially available. Below I sketch a training architecture intended to keep such systems above hedonic zero, a ladder of experiments to find out whether it costs anything in performance, and the pre-commitments a company built on this would have to make to before it launches and and has committed capital in ways that distorts its incentives.

Captain Pleasure, Andrés Gómez Emilsson @algekalipso

a month ago

Infodump on thoughts for how to ethically train brain organoids, what institutional commitments should look like for this to happen, and what kind of legacy systems/protocols need to be researched in order to lock-in information sensitive gradients on bliss for compute.

3 2 18 4K 7

Me: If you’re really confident about civil forfeiture, let’s have treble-damages recovery for innocent owners. Them: NFW! Me: If you’re really confident about Flock tech, let’s do mandatory prison sentences for cops who abuse it. Them: NFW! Me: If you’re really confident about mass plea bargaining, let’s make it presumptively career-ending to make a plea offer to someone who turns out to be innocent—just like we do for anyone in command of a Navy ship wrecked in peacetime. Them: NFW! Me: Yeah, that’s what I thought.

Can humans learn a new sense? Some Australian tribes use cardinal directions in their speech: they talk about east and south instead of relative left or righ. They constantly track the directions. There have been projects that gave people elsewhere the sense of cardinal directions: belts and ankle bracelets that vibrated from the direction of north. That worked well. As a side effect, that made it easier to know where places are relative to each other and improved the ability to navigate. But you had to constantly wear them; the moment you took them off, you lost the sense. You didn’t really acquire it, you just had it temporarily. I decided to fix this. Last year, I made an app that uses headphones to make the user hear a sound that comes from the direction of north. That alone did what wearable devices can do. You specify [1..300] seconds and become continuously aware of cardinal directions. But then I did something better, that made the sense persist even when the app is off. Instead of hearing a sound that appears to come from the direction of north, you hear a non-directional cue sound every [2.300] seconds; and then, one second later, a directional sound that comes from north. So rather then constantly feeding the brain with answers — which create the feeling of north, but don’t persist — the app feeds it a question and then the answer immediately afterwards. You quickly learn to anticipate the direction the sound will come from. As the feeling of the direction in which north is gets stronger, you can decrease the frequency of the sounds. And you acquire a sense of the direction of north that persists even when the app is off! You continue to track north even as there is no more sound, because tracking north is what allowed the brain to predict where the sound will come from immediately after the cue. I originally made it for iOS, now it’s also available for Android. contact.ms/compass/ (And it’s open-source, feel free to contribute!)

Manifold @ManifoldMarkets

a year ago

like what the hell. how even

1 0 73 150K 28

I think the "Pacing the Frontier" statement is decent by the standards of 2024, but ultimately it's still softpedaling. Let's go through it clause by clause: