pumpkin-eater's recent activity
-
Comment on Google releases Gemma 4 in ~comp
-
Comment on Google releases Gemma 4 in ~comp
pumpkin-eater Link ParentSure... firstly, if anybody is interested do DM me for access (the main pitch is privacy by avoiding cloud models, though, so whether there's any benefit to you in having your...- Exemplary
Sure... firstly, if anybody is interested do DM me for access (the main pitch is privacy by avoiding cloud models, though, so whether there's any benefit to you in having your transcripts/summaries living remotely on my server is another question... I am hoping to release it eventually as a free self-hostable project)
It's a website and discord bot that records separate audio files for each participant in a call, clips up the recordings based on the start+end time (it starts recording as soon as everybody joins the chat, but you can DM it to cut off the start of a call, so that general opening banter isn't transcribed), then runs each recording file separately through whisperx (distil-large-v3.5, silero vad, then WAV2VEC2_ASR_LARGE_LV60K_960H to better-align to timestamps). It constructs a transcript, runs through a fixed pipeline to clean up noise words, produces a transcript that uses character names, then runs it through chunked summarisation with qwen2.5:14b using structured output (basically asking it to produce headings for sections and points within the sections).
The LLM is being used both to summarise but also to remove off-topic / meta-discussion. It does an OK job, but my tests feeding transcripts into Opus are wildly better.
At the end, after somebody has signed off on the summary, I turn the combined JSON summary into markdown and send it to Gemini Flash for summarisation.
I've tried Qwen3, Phi4, etc. but struggled to find something better than Qwen2.5 (I'm limited to a 12GB 4070). The whisperx transcript has a bunch of errors (some of which may be cleared up if I ran a combined audio file through, but I don't like that I lose the completely accurate diarisation I get from having separate audio tracks for the different players). I've tried Parakeet too, but accuracy isn't great. Not sure if the problem is the separate audio tracks losing conversational context, accents (mix of different nationalities), or perhaps simply poor diction.
At its core it's a pretty straight-forward transcribe+summarise process coordinated by typescript, complicated by the need to be local-only and my network setup with a linux server but a windows machine that has an RTX 4070 I use for CUDA (my groups don't want what they say to be sent to a cloud model, so that they don't have to worry about an off-colour joke coming back to bite them in the future).
I'm not sure if it's generally useful, but I'm certainly happy to share/collaborate if there's anything there that sounds like it might be useful.
The prompt that I've worked out that gives me the best results is:
You are summarizing a TTRPG session transcript. System: (system name) Background: (1 sentence context for game) Player Characters: - (character name) - (1 sentence summary) Instructions: - Extract key narrative sections from this transcript chunk - Focus on in-game events, interesting character actions, story developments. If in doubt, assume the problems and achievements are interesting, err on the side of completeness - The GM plays all NPCs and describes scenes and results - Players are identified by their character name in the transcript - Describe the interesting things that CHARACTERS DO, places and problems they encounter, what NPCS DO THAT EFFECT CHARACTERS, and how they solve them - USE CHARACTER/NPC NAMES - Only report CONCRETE events that actually happened, not implications - Filter out off-topic chatter, technical issues, meta-discussion.(the biggest problem I've had with prompting is trying to block the tendency of models to say fluff like "raising questions about", or "make deep discoveries", which I don't want them to do, or even worse to summarise moods "after a feverishly pitched battle, the exhausted group..." - I want a simple digest of who did what to whom)
-
Comment on Google releases Gemma 4 in ~comp
pumpkin-eater LinkPerhaps my use-case is too specific, but I've struggled to find a local model better at factual summarisation of transcripts than qwen2.5:14b. So many of the models I've tried try to hide their...Perhaps my use-case is too specific, but I've struggled to find a local model better at factual summarisation of transcripts than qwen2.5:14b. So many of the models I've tried try to hide their lack of understanding with broad statements or implications about what's going on rather than stating concrete facts as requested in the prompt. gemma4:26b (e4b) is the same so far (I'm trying to provide a blow-by-blow of TTRPG sessions)
-
Comment on Nvidia CEO declares AI could start, grow, and run a successful technology company worth more than a billion dollars—excerpt from Lex Fridman Podcast in ~tech
pumpkin-eater Link ParentI don't see why you're twisting yourself around this... what do you mean by "whether to count this one in 2026 predictions"? What prediction? That AGI is achieved, or that somebody in the industry...I don't see why you're twisting yourself around this... what do you mean by "whether to count this one in 2026 predictions"? What prediction? That AGI is achieved, or that somebody in the industry will claim AGI? If the former, I don't see how this has any bearing on the matter (especially as Kacey points out, given Jensen stands to profit from such a claim being made, even if untrue). If the latter I'm pretty sure there have been random hypemen claiming AGI for a while now...?
The most generous reading I have of the interaction you quoted from the podcast is "nVidia CEO says CEOs can be replaced by LLMs": Friedman says: "an AI system that’s able to essentially do your job. ... start, grow, and run a successful technology company" (emphasis mine).
-
Comment on René Redzepi, the head chef and co-founder of Noma, announced he was resigning from his internationally acclaimed Copenhagen restaurant following allegations that he had physically abused his staff in ~food
pumpkin-eater Link ParentYour feeling is backed-up, I think, by the language he used in his weird documentary-style announcement to staff where he said they'd still see him around just not in the same way, and he was...Your feeling is backed-up, I think, by the language he used in his weird documentary-style announcement to staff where he said they'd still see him around just not in the same way, and he was going to plan the next chapter.
-
Comment on Can coding agents relicense open source through a “clean room” implementation of code? in ~comp
pumpkin-eater (edited )Link ParentMissing context, perhaps: it was a wholly client-side javascript application so the source code was available to the LLM to copy. The fact that the datastructures were copied is what tells me the...Missing context, perhaps: it was a wholly client-side javascript application so the source code was available to the LLM to copy. The fact that the datastructures were copied is what tells me the source code was used.
Edit: for clarity, I'm not claiming this is identical to what is happening with chardet, but I think these are all echoes of the same fundamental issue of LLMs and copyright-washing (first in training, now in application). If these cases are ruled as legal, what does it mean for the whole concept of copyright?
I think doing this to gratis/libre software is especially galling because we're really not asking for much. Often we're simply asking for people to credit us with our work, or to pay forward the benefits they got back/forward by sharing the improvements they've built on top of our work.
-
Comment on Can coding agents relicense open source through a “clean room” implementation of code? in ~comp
pumpkin-eater LinkThis whole subject really depresses me, especially because it hits close to home: somebody used AI tools to clone one of my free projects after I refused their demand to open source it under...- Exemplary
This whole subject really depresses me, especially because it hits close to home: somebody used AI tools to clone one of my free projects after I refused their demand to open source it under threat of AI cloning (they wanted to put it into their site as if they'd created it, rather than linking to my site - I knew they'd do the exact same thing if I open sourced it, which I had originally planned before realising that's exactly what they'd do, and cut me out other than a source of free labour).
Their clone was comprehensive, all the way down to my file format, the features they said were pointless and wanted me to remove, and the in-application Help features. But they have a bigger audience. To make matters worse, they also had the gall to credit me as the developer of the "initial prototype".
Given the courts seem unlikely to come down against LLM companies training on material they don't have rights to use, I'm not holding my breath that this sort of license-washing will fare any better.
-
Comment on Tell me about your favourite web-based logic puzzles! in ~games
pumpkin-eater (edited )Link ParentThis is really fun, thanks for posting it! I've tried to come up with ideas for word games but was stumped and started to think all the ideas had already been made into games, it's great to know...This is really fun, thanks for posting it! I've tried to come up with ideas for word games but was stumped and started to think all the ideas had already been made into games, it's great to know that other people don't have this problem :-)
What I couldn't figure out is whether it's a good thing to get a high score, or a bad thing (i.e. does getting a low score actually mean you're more creative with your language instinctually? It'd be fun to play the game in reverse)
P.S. I think the answers should be shown immediately... I suspected (and confirmed) that LLMs are very good at the game, so if somebody's wanting to cheat that's probably the easiest way to do it
-
Comment on Ian McKellen does relevant Shakespeare on Stephen Colbert in ~society
pumpkin-eater Link ParentGreat find! Here it is on YouTube for people who avoid Meta: https://www.youtube.com/watch?v=bsbW-3pTvhAGreat find! Here it is on YouTube for people who avoid Meta: https://www.youtube.com/watch?v=bsbW-3pTvhA
-
Comment on USB-C PD all the things! in ~tech
pumpkin-eater LinkThis is a really cool idea, but the interface seems like a mistake, there's no way to physically see what voltage a given adapter is going to deliver unless you plug in a meter. You could make a...This is a really cool idea, but the interface seems like a mistake, there's no way to physically see what voltage a given adapter is going to deliver unless you plug in a meter. You could make a simple mistake and push 24V when you thought it was configured for 12V... dip switches seem like they'd be a better approach here, no?
-
Comment on Is there a service that a regular joe can use to provide subs or dubbing for a movie? in ~movies
pumpkin-eater Link ParentIf it's content you hold rights to, yes. Expect something on the order of £3000/hour. As far as I know, Voquent will do one-offs (I'm tangentially involved in the space, never worked with Voquent...If it's content you hold rights to, yes. Expect something on the order of £3000/hour.
As far as I know, Voquent will do one-offs (I'm tangentially involved in the space, never worked with Voquent - I do automation of project submissions for dubs & subs for clients with pre-existing commercial agreements with the translation houses)
-
Comment on What ridiculous thing would you spend billions on? in ~talk
pumpkin-eater LinkA personal ISS with a spin gravity section, a big observation bubble, a telescope+camera, and a lifetime ticket for free rocket journeys for me & my immediate family to it. Also: R&D on a compact...A personal ISS with a spin gravity section, a big observation bubble, a telescope+camera, and a lifetime ticket for free rocket journeys for me & my immediate family to it.
Also: R&D on a compact all-terrain poo-locating-and-scooping robot for our dogs
-
Comment on What ridiculous thing would you spend billions on? in ~talk
pumpkin-eater Link ParentIf they actually believe what they're wishing for, it seems awfully philanthropic...If they actually believe what they're wishing for, it seems awfully philanthropic...
-
Comment on Tilde is kill? in ~tildes
pumpkin-eater Link ParentWhat I do with DNS challenge is run cert renewal scripts via cron on a different machine, and then rsync the certs over to the webserversWhat I do with DNS challenge is run cert renewal scripts via cron on a different machine, and then rsync the certs over to the webservers
-
Comment on ‘Don’t ever assume there’s anything to eat!’ Twenty-nine tips for perfect vegan holidays, from where to go to how to order. in ~food
pumpkin-eater Link ParentJainism takes more of this approach of respecting all life rather than just animal life, worrying about the impacts of their actions (e.g. about harming insects in the soil, even avoiding killing...Jainism takes more of this approach of respecting all life rather than just animal life, worrying about the impacts of their actions (e.g. about harming insects in the soil, even avoiding killing entire plants). I've always thought it a fascinating philosophy.
-
Comment on Scientists estimate European heatwave caused 2,300 deaths last week in ~enviro
pumpkin-eater (edited )Link ParentThis is just standard epidemiology... these sorts of events cause both excess mortality (people who die now who would otherwise have died much later) and displaced mortality (moving forward deaths...This is just standard epidemiology... these sorts of events cause both excess mortality (people who die now who would otherwise have died much later) and displaced mortality (moving forward deaths that would otherwise have happened in the near future).
It's bad (useless, really) reporting not to go into detail. The answer is air conditioning (long-term it's emissions controls, but we all know how that's going...). How do you overcome our European bias against AC? Show people that they are in danger from these events. If you don't give people the information they'll just handwave it away and not take any action.
-
Comment on Scientists estimate European heatwave caused 2,300 deaths last week in ~enviro
pumpkin-eater (edited )Link ParentI think that's a very uncharitable interpretation of what I said. What exactly makes you think I am wanting to shrug off deaths? edit to expand on this: Reuters is quoting scientists, but the...I think that's a very uncharitable interpretation of what I said. What exactly makes you think I am wanting to shrug off deaths?
edit to expand on this: Reuters is quoting scientists, but the information in the story is abstract and limited: in particular it's not actionable. Everybody knows that heat waves kill people. The models the scientists are using will take into account who is most at risk. Those people need to see that so that they have a reason to buy air conditioning units that overcomes their bias against it.
-
Comment on Scientists estimate European heatwave caused 2,300 deaths last week in ~enviro
pumpkin-eater LinkWhat isn't clear from this reporting is how many of these are excess deaths, vs just moving forward deaths that would have happened in the next few weeks/months anywayWhat isn't clear from this reporting is how many of these are excess deaths, vs just moving forward deaths that would have happened in the next few weeks/months anyway
-
Comment on Apple overhauls EU App Store rules following penalty in ~tech
pumpkin-eater (edited )LinkAs a heavy Apple user for most of my life, my view of the company has significantly changed because of this - the level of constant petty malicious compliance makes them look weak (and tasteless)As a heavy Apple user for most of my life, my view of the company has significantly changed because of this - the level of constant petty malicious compliance makes them look weak (and tasteless)
-
Comment on What are your favorite vegan pre-packaged foods? in ~food
pumpkin-eater LinkIn the UK/Ireland, Marks & Spencer do a really good vegan "no chicken kyivs", they've got a really nice texture and a fantastic garlic flavour - I prefer them to chicken kyiv.In the UK/Ireland, Marks & Spencer do a really good vegan "no chicken kyivs", they've got a really nice texture and a fantastic garlic flavour - I prefer them to chicken kyiv.
That's really helpful advice thank you! And encouraging to hear you think it's interesting... I keep doubting myself whether it's a worthwhile side-project to continue with (especially without buying a new card with more vram, which would be way over budget for a side project, this is really my only side-project that uses local inference so hard to justify)
I haven't tried giving examples, tbh, I've been using quite small context windows for chunking and thought any sort of example would blow the context out, but you make a good point... I'm so used to the frontier models broad zero-shot capabilities. I'd originally imagined you'd need a representative example of the type of conversation in the chat, but maybe I'm presupposing what a solution looks like... I should try coming up with short examples.
On whisper recognition errors: I'm not sure if my problems are broadly applicable, speech recognition tends to have problems with celtic accents, and on top of that we go off on tangents a lot, we aren't disciplined roleplayers. We're a majority Northern Irish group, and the American members are so used to us that we all speak at full speed everyday speech (just checked and the effective WPM after noise word filtering is 1.5x for Irish vs American players). Another part of our broader group has German and Afrikaans accents, they've just started their game and I'm really interested to find out how it'll cope with that.
The more broadly applicable problem I did have, originally, was reassembling transcripts given that I split up recognition per-speaker - as you probably know whisper timecodes drift in longer recordings, whisperx with wav2vec alignment was a big help.
LLMs do seem good at seeing through the recognition errors to get a general sense of what's going on (I assume because of how redundant English is).
One thing I've been wanting to try is an audio model, to try blending text and audio tokens, i.e. a stream of "[speaker name] [audio tokens]" and see if it can do a better job at transcription/direct summarisation with more conversational context available.
One experiment I tried originally was augmenting transcripts with structure prior to summarisation, but it didn't really pan out (unclear whether that was my prompting, or a flawed approach): my thought was that TTRPG sessions are a collection of scenes, where the GM establishes a scene, PCs and NPCs interact, and then the scene is resolved.
My approach is a limiting factor with experiments, though, because one of the privacy goals I set is that transcripts get thrown away within 7 days (keeping only the summaries) so I don't have a large backlog of sessions to run experiments against. I wanted to do that to make sure that players didn't feel the need to self-censor with the idea that their words would live on in a database forever, and I think it's worthwhile, if the sessions being recorded changed how people behaved I think I'd stop the project, because it's supposed to be fun escapism with friends
(apologies for the overly long response...)