martind @dekstop.bsky.social · 12/08/2026Turns out a contemporary phone with zoom lens and manual exposure does just fine! 000
martind @dekstop.bsky.social · 29/07/2026TFW Google Meet has all the novelty filters but no adjustments for basic lighting, colour correction, or removing the distracting visual glare from my eye glasses 100
martind @dekstop.bsky.social · 26/06/2026This week's vibe coded project: an AU synth plugin that loads C++ "scripts" to generate sound, with automatic recompiles whenever the source file changes (intended as an audio dev tool in size coding projects, could feasibly also become a basis for a C++ live coding setup) github.com/dekstop/GenS... 010
martind @dekstop.bsky.social · 14/06/2026Accidental achievement unlocked: got butchered in Nova 2026 intro compo for flippant language about my AI use. Oops 100
martind @dekstop.bsky.social · 23/05/2026Just caught myself working on two separate projects in parallel for the last couple of hours, presumably it felt like a natural way to bridge the prompt iteration waiting time. 000
martind @dekstop.bsky.social · 13/03/2026Deal of the century. Should I have taken more than the five packs 000
martind @dekstop.bsky.social · 17/09/2025Our publisher @plutopress.bsky.social surveyed their authors whether their published work should be licensed to train large language models. 15% responded "Yes", 41% responded "No", 30% responded "Maybe", depending on how the data is used. Full report: www.plutobooks.com/wp-content/u... 110
martind @dekstop.bsky.social · 13/09/2025(Caption: screen recording of a chat session with the prompt: "say the time". The LLM abides: it requests to run the "say" shell command with an inline call to "date". User approves, computer reads out the time using MacOS speech synthesis.) 000
martind @dekstop.bsky.social · 12/09/2025Playing with fire: my local LLMs can now run web searches, read local files, execute Python code, and run shell commands (in LM Studio with local MCP server) 120
martind @dekstop.bsky.social · 07/07/2025For reference, here the original schematic before I asked it for some corrections 000
martind @dekstop.bsky.social · 07/07/2025So happy, ChatGPT helped me design a bipolar power supply, the schematic looks very professional. And so efficient -- who knew that a single line could supply both the positive and negative rail! 110
martind @dekstop.bsky.social · 22/04/2025Many good sections, but take particular note of the simple conclusion at the end! 000
martind @dekstop.bsky.social · 05/11/2024Have officially begun the Charlie Brown Christmas season. The instrumental version for now because I'm not a loon 020
martind @dekstop.bsky.social · 09/08/2024Exceptional discussion in the FT of the mechanics of riots, drawing from decades of social research incl 2011 riots in the UK archive.ph/PONGF 000
martind @dekstop.bsky.social · 29/07/2024My still on Twitter / not quite actually on Twitter conundrum has just resolved itself, my account failed their new mandatory age verification. Finally free! 000
martind @dekstop.bsky.social · 17/04/2024Is AI even worth it? @molly.wiki shares her observations so far www.citationneeded.news/ai-isnt-usel... 000
martind @dekstop.bsky.social · 14/04/2024An AI UX paradigm -- software that writes itself in response to user input? 010
martind @dekstop.bsky.social · 10/04/2024PS here its initial answer before I added brackets to avoid an apparent confusion about operator precedence. I suppose this one is at least arguable, if maybe a cop-out 000
martind @dekstop.bsky.social · 10/04/2024Here a recently released state of the art LLM that claims to beat GPT-4. (Once more, the answer is incorrect.) 100
martind @dekstop.bsky.social · 07/04/2024Built an AI to help me navigate my vinyl collection 🤖 Now how can I get it to do my dishes (It's RAG over my Discogs collection plus MusicBrainz metadata. An embeddings lookup plus GPT for filtering/enrichment.) 000
martind @dekstop.bsky.social · 01/04/2024GPT4 reportedly executes Python to evaluate mathematical expressions, but most LLMs don't (yet). I've found the following expression useful to test if a model can do maths -- none of the results shown here are correct. 200
martind @dekstop.bsky.social · 15/03/2024TLDR: Current-generation LLM based AI is best thought of as an assistive technology. It can improve the efficiency of an analyst. But it can only enhance human analytical thought, it cannot replace it.* * ... except maybe in scenarios where the cost of a wrong judgment is tolerably low? 101
martind @dekstop.bsky.social · 15/03/2024I also wonder if this might nudge current gen AI closer to the trough of disillusionment of the Gartner hype cycle: towards the disappointing realisation that human involvement and specialist knowledge is still needed to produce meaningful outputs. 100
martind @dekstop.bsky.social · 15/03/2024I wonder if this will be one of the hard limits of language model based AI, and whether it might take another indeterminate period for newer models to push through that. 100
martind @dekstop.bsky.social · 15/03/2024Further, it is easy to forget how much human oversight is needed to assess and correct the outputs of generative AI. Although outputs can look persuasive, they simply reproduce patterns encountered in the training corpus. Current LLM based AI cannot actually reason about the data you give it. 100
martind @dekstop.bsky.social · 15/03/2024(LLMs also won't get the data for you, look at the data for you, generate the map for you. And chances are you will need to tweak the generated python code to make it work. Some but not all of these gaps can be addressed with task specific tools such as copilot.) 100
martind @dekstop.bsky.social · 15/03/2024You can ask a LLM to make a map, and it will happily generate the python code. But it won't review the resulting distribution, decide on an appropriate visualisation style, overlay other indicators, run a regression model, compare with a prior expectation, or take other steps an analyst might. 100
martind @dekstop.bsky.social · 15/03/2024LLM-based AI can help with the technical implementation of specific procedures, writing up findings, brainstorming approaches. But as far as I have seen it still requires a specialist to guide it, to ask the right questions and make the right judgment calls. 100
martind @dekstop.bsky.social · 15/03/2024Many professionals seem to be under the wrong impression that current-generation AI can analyse data for you. I'm happy to be corrected, but I looked into it and from what I've seen it absolutely can't. Large language models can however be useful in assisting a human analyst. 100
martind @dekstop.bsky.social · 08/03/2024Recreating my avatar in AI. Infuriatingly it's always using the same default man face no matter the prompt :cry: 000
martind @dekstop.bsky.social · 08/03/2024The sequinthropocene: a vision of a future where everything is covered in sequin waste. (Coined by a friend.) 000
martind @dekstop.bsky.social · 08/03/2024Happy international women's day! I just spent too long wrangling Bing Chat into avoiding all the bad clichés of gendered representation, this is one of the less bad ones, it still ignored the request to avoid pink. I remain sceptical about our impending AI future 😅 #IWD #IWD2024 000
![Reddit screenshot: "Mod note: we are banning AI 'Neural Howlround' posters."
Text summary: "in the interest of full transparency, I just wanted to mention that we also (quietly) ban a bunch of schizoposters and AI 'Neural Howlround' posters, under the "spam" rule, since the contents of the posts are often nonsensical and irrelevant to actual AI. [...] TLDR: LLMs today are ego-reinforcing glazing-machines that reinforce unstable and narcissistic personalities to convince them that they've made some sort of incredible discovery or created a god or become a god. And there is a lot more crazy people than people realise. And AI is rizzing them up in a very unhealthy way at the moment."](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:e2xtlgh76mdcucctdq24xjxb/bafkreiea2rbvddkypzwuyilso6v4k3lfwxvtt2qphd7b72walee2rgknci)
![User: Summarise this conversation in a compact symbolic form that an llm can interpret to recall the full content. Don't bother including human readable text, focus on llm interpretability only
ChatGPT: TASK[CompareOpenLLMvsGPT4];SOURCE[HF/OpenLLMLeaderboard/OfficialOnly];METRICS[IFEval,BBH,MATH,GPQA,MMLUPro];TARGET[GPT-4(2023)];MODELS[Yi-1.5-6B(16.7B),Yi-1.5-34B(Base,32K,Chat,Chat-16K)];ARCH[Dense,LLaMA];TUNING[Base,Chat,RLHF];SIZEBRACKETS[<10B=Fail,10–30B=Mid,30B+=Best];FINDING[NoModelMeetsGPT4,TopModel=Yi-1.5-34B-Chat@BBH60.67%,GPQA13.06%,MMLU39.12%];TREND[Size+RLHF=BetterReasoning,TradeoffsExist];](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:e2xtlgh76mdcucctdq24xjxb/bafkreibamum4j243sv44uvaxyl5ejxbusve3yc5skzlwdjyivbcifxacvy)
