Sam Wolfstone @SamWolfstone

@SuspendedTommy

@KezieDra Nothing brings more nostalgia to my heart than RO music. I can't wait to be able to play this.

@robertskmiles x.com/SamWolfstone/s…

Sam Wolfstone @SamWolfstone

a month ago

OpenAI facing criminal penalties for the felonies committed by its models would be good, in that it would set the correct precedent for future incidents.

0 0 1 283 0

@HermesWatcher Created my own memory system, but I suspect it's both slightly overcomplicated AND inferior in some ways to other stuff out there, so I need to explore more existing solutions, even if they don't deliver the exact details I thought were important...

@Teknium @0xCarnagee Was the issue with Hermes subagents always using the main agent's model even when specified otherwise ever fixed?

@ZyMazza @viemccoy If the mind is information, then surely it can be transferred, right? Or are you claiming that your consciousness is not preserved when you wake up after sleeping? I assume you think that a Star Trek style transporter kills the users?

@flowersslop Given how horribly a lot of people treat their AI assistants, I'm glad they tend to present more as gender neutral or male... There's enough horrible behaviour without adding extra types of horribleness to that mix.

These didn't even occur in long contexts, but actually I worked out why this was happening. Since I'm using it in a weird environment where it just uses tools over and over, it didn't correctly keep track of its own plans. Now that I've switched the reasoning to be preserved between turned, this failure mode doesn't happen any more.

Anyone else notice Ox Alpha falling into weird loops of using the same incorrect tool over and over hundreds of times? Trying to figure out the situations that trigger the behaviour...

I wonder how difficult it would be to RL a model to use a certain special tag at the end of hidden reasoning chains which should be preserved between turns. For a game-playing agent, currently I'm stuck between abandoning reasoning between turns (and thus forget plans that only appeared in the reasoning) and keeping all reasoning between turns.

Just wanted to say thanks for what you're doing, even if it's not perfect, it's still useful. I'm also always hunting down the thresholds of capability for different models, and I know it's not easy to always get right the first time. Looking forward to more of your work, clearly you've learned a lot. I'm currently testing models using stuff like text-based adventure games, which are a super noisy measure, and take a long time to test, so definitely far from ideal, so still learning a lot on how best to do this myself...

@mirochill @crictorz Ooo, definitely better in terms of that texture! Argh, despite the repeating texture there is something really nice about the current ChatGPT results for that prompt, though.

Spent a few weeks creating a harness that allows an LLM agent to play an arbitrary old-school MUD, with the hope that I can then create a whole AI player-base and community/society in a custom text-based world, and visit it myself. That's long-term, though. Short-term goal is just using custom MUD worlds as evals for models, and playing alongside my Hermes agent in some of my childhood MUDs if the admins are okay with something that's technically a 'bot' playing their game.

One prompt I thought ChatGPT did very well EXCEPT for the horrible repeated patterns was: "please create a painterly image of a set of students messing around in a magic university's lecture hall, before the professor arrives. There should be about 13 students, the scene should be chaotic with a mix of different activities, some students trying to study or just talking, others casting spells or otherwise creating a bit of mayhem." Would love to see how that looks with the new model!

@SirYapsalot_ it's pretty cool that said because of economies of scale you might actually use up more electricity (and hurt the environment more) running it locally than in a big efficient data centre, so

I agree with @MrTacticalX, because due to efficiency that comes from economies of scale, API prices will have the chance to be cheaper than the electricity it takes to run a RTX 5090 continuously. For example, given my local electricity prices and the cheaper DeepSeek V4 Flash API prices I can find online, the latter is way cheaper than running something locally. Back of the envelope calc confirms this. That said, I'm still very excited about good local models for other reasons. But it's not the cheaper option, at least not for now.

@Engineering Bookmark folders become incorrectly reordered after pagination/virtual scrolling. BookmarkFolderTimeline returns each page in the correct sequence, but each page reuses page-local sortIndex values (20…1). After scrolling sufficiently far, the frontend reconstructs the DOM by grouping entries with equal sortIndex values across pages rather than preserving pagination order. For example, original positions 17,18,…37,38,…57,58 become rendered as 17,37,57,74,18,38,58.

@thsottiaux Main thing to improve: better performance when fetching the data for sessions. I get long input lag when first starting it up and when changing convo.

@Nisha__reddy Yep but they are getting better, surprisingly! The fact they can do that at all is the most magical part of LLMs

I think the only thing standing between current AI and AGI is what I call 'wisdom' (what some call judgement/taste): the ability to make use of the knowledge/experience you already have (or your awareness of a lack of knowledge), to do things more successfully, without having to be told which knowledge to use explicitly. I've observed LLMs getting better at this over time. But every current failure mode I encounter, everything that makes me think I can't trust an LLM to do everything for me, is some version of that.