May-2026

A thought stream from May 2026.

Mon 04-05-2026

• my-mindware
  • I’ve just incorporated a transcription software called WhisPaste into my mindware workflow which is so far really helpful because it allows me to just dictate thoughts and automatically paste them wherever my cursor is. This thought itself was typed via this dictation method.
  • Something I’d like to figure out is whether I should indicate whether a particular thought has been dictated versus typed. But I’m starting to think that this might not be necessary because I can just edit the thoughts retrospectively via manual typing and the dictation is primarily a method of physically capturing thoughts more quickly/fluidly/efficiently.
  • A thing I would like to experiment with is having a log section in each task/project note (i.e., note representing a particular task/project I’m working on) where I can record thoughts related to that task/project and then later synthesise them into a collection of learnings that emerged from completing that particular task/project.
• GenAI
  • I’ll use this tag to refer to the kind of AI systems we typically associate with the term “AI” (i.e., LLMs and any systems/tools built with LLMs), and will reserve the term/note AI for a more general conception of artificial intelligence (e.g. recommendation systems, autonomous vehicles, symbolic AI systems - though perhaps these do/could incorporate LLMs as well…). Anyway, it’s an initial distinction that can be refined later.
  • Addressing duplication of content / generative efforts.
  • I had a thought just now about the fact that many minds will have similar things they’re interested in generating outputs for that they prompt LLMs for, meaning there’s probably a lot of duplication in content that’s generated and also probably a lot of content generated that other minds would find useful that they may not necessarily think about prompting for. However, it’s also important to recognise that there may be nuances in the content requested based on certain contextual factors about the prompting mind, stuff specific to their situation/needs, etc., and so in these cases, it’s useful/necessary to just generate new content. I think it really depends on the resource usage of generation, because if e.g. generating a report about a particular item of interest takes a lot of computing power and it uses a lot of energy and other resources then it’s important to consider how we can minimise this duplication of generation by e.g. creating a database/archive of all generated content that can be queried, and then you can quickly see if there’s something relevant to what you’re interested in that’s already been generated; this would also save you prompting (so, also addresses duplication of human prompting efforts - though adds the effort of searching), and perhaps the mind that has already generated content relevant to you has created a more comprehensive prompt than you would have done and considered things you haven’t thought of. Also, you can see if the content you requested isn’t in the publicly available database and in that case it’s fine to go ahead and generate it - there will be no duplication. We could all be tapped/connected into this public database and we can choose which of our outputs to make publicly available and which to keep private.

Sun 03-05-2026

• mindware
  • I’ve been watching this video: (2714) How I use LLMs - YouTube and want to extract some of these ideas on how to use LLMs, and any suggestions for functionality of the future.
    • Co-reading with LLM
      • You can paste an e.g. chapter of an online book, or academic paper, into an LLM and ask it to summarise, and then clarify any points of confusion as you go along.
      • It would be good to be able to e.g. highlight passages in a book either on computer or e.g. kindle and ask questions directly, rather than flipping back and forth between the LLM chat and e.g. book PDF.
    • Embedded tool use within LLMs
      • As of when this video was recorded, ChatGPT and Claude call on programming tools to solve harder problems (for example, the multiplication 14923743453 * 342439248392) that the normal probabilistic predictive generation would likely hallucinate the solutions to (e.g. Grok and Gemini, which did not outsource to programming tools).
      • However, in either case, the LLM will output an answer to the question as though it is correct. Which made me em that it would be good if LLMs were able to indicate some level of “epistemic uncertainty” in whatever answers they generate - so, they have an awareness of when they lean heavily on next-token-prediction as opposed to more deterministic, “failsafe” tools like e.g. ChatGPT’s use of Python to calculate the value of a complex expression. So, if they used such a tool, or e.g. were able to directly quote/draw upon a particular source, they should indicate this (as e.g. “low epistemic uncertainty” / “high accuracy”), as opposed to when they provide a “gist” response based on all possible/feasible responses (to which they should append a “high(er) epistemic uncertainty” disclaimer).
      • I’m not sure whether this kind of capability already exists in some capacity in current models / is in development pipelines, but I’m not aware of any major developments (as I still don’t fully trust output when it comes to e.g. appropriately/accurately citing/quoting sources).
  • Auto-summarisation of AI chats
    • A lot of learning and deep exploration of ideas/concepts happens within a singular LLM chat, and I’d like a way to capture that. One way of doing this could be to e.g. define a “chat completion” where you indicate that you’ve finished with a particular chat, which closes and archives it, and automatically generates a summary of what you discussed, what ideas were presented/covered, how your understanding seemed to develop. Then, perhaps you can also add your own reflections to this summary, and the result is appended to a running log of conversations/learnings.
    • There could be a general serial stream which all chat summaries are appended to, and then each summary could also be tagged with any related objects and/or appended to the raw thought sections of these objects (well, their notes).
    • Then, if you want to return to something that was discussed in a previous chat, you can just copy-paste the relevant summary into a new chat.
    • This would also save (free-up) tokens / space in the context window ? and could be a convenient way to e.g. enact a cycle of continually condensing and expanding objects of discussion: you discss something with the LLM (expand on it), the summary condenses it, you later expand uupon this summary, and these expansions are later condensed into a new summary, etc. - a continuous oscillation between elaboration and compression (diverge-converge).
• my-mindware
  • I’m trialling an app called “Offscript” (I paid for a month subscription which was <£3) mainly to learn what features/functionality I find useful, and would like to emulate/incorporate for my own mindware.
    • The app allows for near-instantaneous transcription of voice recordings, each recording stored/presented as a time-stamped voice note, with the ability to assign tags, merge notes, and export/share in different formats.
  • In terms of generating personalised inputs, I tried Gemini’s “deep research” functionality for the first time, where I pasted in a bunch of thoughts about collective-intelligence and areas related to it I’m interested in learning more about, and then asked it to generate a comprehensive report on all of it. I didn’t really know what to expect, but from what I’ve read so far, I’m happy with it and looking forward to delving deeper.
    • However, I did notice that it made claims using specific wording I’d used in my prompt as though they were gleaned from the sources it examined, which made it a bit tricky to tell apart my own emergent thoughts versus the emergent/encountered thoughts of other minds.
    • So in future, I think I’ll give less “my-worldview-flavoured” prompts, and then will hopefully get a more representative overview of existing thoughts about stuff (that aren’t tainted with how I’ve framed / talked about them in the prompt).
  • I also discovered “Piper TTS” - a free, open-source interface for converting from text to speech with loads of voice options, and ability to download the result.
    • I would like to try using this to generate custom auditory inputs I can listen to while humaning (doing general life stuff, like commuting, washing up, etc.).
  • I also discovered the FUTO voice input and keyboard apps, which are, acu (as currently understood), alternatives to e.g. Google speech ? and Google keyboard, which I learned record all data from e.g. what you type/say, which isn’t ideal (although privacy concerns are currently not a major priority for me, but this is something I need to explore and think more about).
    • So, I now use these FUTO alternatives and like them so far - I especially like the ability to e.g. just open an Obsidian note and start writing in it with my voice.
  • Using audio transcripts
    • I asked Claude to create a Python script for merging all my transcripts from all audio recordings in a particular folder, into a singular text file, where each original transcript is separated by a timestamp.
    • So now, I’d like to play around with the resulting merged transcript and see what I can do by querying LLMs on it. Here’s a starting prompt I’ll try:
      • Over the past few days, I’ve been making ~20-30 minute voice recordings, just sharing thoughts about various things I’m interested in thinking more about and developing ideas around. The recordings have been transcribed and merged into a singular text file, where each transcript is separated by a timestamp. I’m now interested in seeing what I could do / what functionality I could extract from feeding the transcripts into an LLM. Some possibilities I would like to explore are: summarising everything discussed, extracting key ideas / recurring themes, generating a list of all distinct ideas, assigning each thought/sentence to relevant notes (e.g. tagging them), identifying points that could do with expansion/clarification, editing the transcripts to improve clarity while preserving raw thoughts, giving suggestions on things to think about in order to expand certain thoughts, generating a sort of “reply” in which you ask questions about the ideas and build upon them, to form part of a broader, unified, ongoing, long-form conversation about everything I’m interested in, etc. Please (1) synthesise/summarise everything discussed in the transcripts and (2) give an overview of which of the functionalities proposed previously would be feasible, the best way to implement these, and general thoughts on how to make best use of LLMs in processing these transcribed thoughts.
      • I asked just Claude and ChatGPT for now, and found the responses useful but cbb (can’t be bothered) to expand on them rn as it’s 22:55 so I’ll resume tomorrow.
• acu tags
  • I’m using this tag to mean “as currently understood” in conjunction with acc (as currently conceived) just to see if any nuances in usage emerge. If not, I’ll merge them into one tag - the one I converge on using most / all of the time.
• cbb tags
  • “can’t be bothered”

Sat 02-05-2026

• my-mindware
  • Screen recording
    • So today, for the first time, I started screen-recording as part of my output attempts (i.e. approaches to externalising thought). I was doing some work on a uni project, and I wanted to talk through what I was doing and explain thought processes.
    • Some of the reasons for doing this kind of thing include:
      • Identifying patterns of thinking and picking up insights about how problems are solved and e.g. the thoughts/approaches with the “greatest contribution” to arriving at a solution/answer to a problem/question.
      • Generating data for training AI systems (personalised, or general) and for feeding into an output pipeline which an AI system can use to identify key features of my thinking (and thinking in general) and things to improve.
      • Organising my own thinking and clarifying my approach; I find that the act of explaining what I’m doing, in the knowing that some other entity - human or otherwise - may end up inputting what I’m outputting, activates my mind in a way that allows me to approach things more effectively than otherwise.
      • And obvs like always, there is a lot of overlap between these reasons, and many more reasons which I can’t be bothered to unpack right now.
  • Generation of (personalised) inputs
    • So, last night, I experimented with asking ChatGPT to generate a “bedtime podcast episode” on the latest AI developments and where things look to be going. It worked well, and I was impressed with how realistic the narration sounded.
    • But, I realised I don’t have the option of increasing the play speed (in ChatGPT), which is important functionality for me. So today, before going on a little walk, I wanted to set up an auditory input which did allow me to increase the play speed. I downloaded a text-to-speech app, “Speech Central”, asked ChatGPT to generate a ~20 min script about a bunch of stuff I’ve been thinking about lately, pasted it into Speech Central, and listened to the “episode” on my walk.
    • Overall, I enjoyed it and found the actual content of the script useful and relevant to me (well, it was useful and relevant by design, as I requested stuff relevant to what I’m thinking about at the moment), and I liked how it would sometimes address me directly; it made me feel like I was involved in some kind of mission briefing thing and it was just fun to feel like the AI was talking to me.
    • However, the voice was very robotic (unlike the ChatGPT voice), so I’ll look at how I can upgrade that, and explore other apps as well.
    • Additionally, I noticed that the script took quite a while to generate, so it would be good to e.g. have stuff regularly generated based on my recent outputs/inputs (e.g. things I’ve been talking about / reading about lately) which I can easily dip in and out of throughout the day. But ofc, need to prioritise research into e.g. environmental burden of such pursuits.
• mindware
  • Enunciating thoughts
    • In terms of how the mindware stuff I mentioned above (that I’m trialling in my own thought endeavours) relates to mindware more broadly, I think this kind of “enunciation of thinking” sts (so to speak) - i.e., explaining and demonstrating thoughts process and resultant actions - will be important for enhancing collective-intelligence, as it will enable us (including AI systems) to better understand how goals are achieved (e.g., the goal of accomplishing a task, like my uni project work) and extract information/knowledge/wisdom (DIKW) about the effectiveness of different approaches and what to avoid (and what to do instead).
    • For example, I would find it useful to have the ability to e.g. feed a bunch of my narrated screen recordings (of me doing a particular task) into an AI system and receive tailored advice on how to sharpen my thinking, ways to improve my approach, actionable advice for completing similar tasks more efficiently/effectively in future, etc.
• tags
  • Until I feel a tag has been sufficiently embedded/used in my outputs, I’ll often state what it stands for / means nearby (e.g. in brackets, as with the use of the sts tag above), just to aid processing/inputting (for both myself and other minds).
• DIKW
  • Synthesise data → info → knowledge → wisdom (via intelligence)
  • How to capture knowledge, intelligence, wisdom?
    • Just think of all the accumulated KIW over a person’s lifetime - how can we extract that throughout a person’s life, and feed it into Mind, overall blob of all human KIW?
  • Actually, though I like the idea/concept/framework of the DIKW pyramid (don’t know where I heard it from - but for now, not important, as I’ve previously discussed the desire to break free from constraints imposed by the necessitation of demonstrating thought origins via e.g. reference/citation - though I wish to expand on this) I also like the em acronym KIWI for e.g. knowledge, information, wisdom, intelligence (though conceptually, really, “information” should appear before “knowledge” as information is a prerequisite/foundation of knowledge ? but whatever, I like KIWI).
• em-enc
  • An example of what I mean by “emergence” is the em tag above, which I used to indicate that the acronym KIWI emerged in my mind independently (i.e., I haven’t encountered it as a thought outputted by other minds) - however, the use of em is to highlight that I recognise this particular acronym for this concept may already exist, and so I’m not trying to claim any originality.
  • In fact, I want to clarify that I’m virtually never making any implicit originality claims, because I understand that a lot of thoughts that emerge in a mind (including mine) probably already exist at least in some capacity. But if we always burden ourselves with the need to identify all currently existing similar/identical thoughts, that just depletes mental (and general) energy that could otherwise be channelled into further developing our own emergent thoughts.
  • I also think it’s useful to label thoughts as either emergent or encountered, because it can give us an idea of e.g. whether a particular thought is emerging spontaneously/simultaneously in many minds, and hence give insight into general truths (because for example, if many minds are independently arriving at a particular conclusion about X at a certain time, we can look at e.g. features of these minds, such as location, social context, media exposure, etc. and determine if there are any illuminating biasing factors - if not, this may suggest there has been a true convergence on a particular truth/aspect of reality).
    • We can also consider things such as the inevitability vs contingency of certain developments. For example, the development of calculus by both Newton and Leibniz: we generally accept that they developed calculus independently, and so this raises a question about whether the introduction of something like calculus into the intellectual zeitgeist (and into the more timeless overall mathematical toolkit) was inevitable, given all prior mathematical advancements.
    • Considering AI, is it the case that any “intelligent civilisation” reaches a threshold of knowledge / technological advancement beyond which the creation of transformative AI is inevitable?
  • Explicitly labelling emergence / classifying thoughts based on emergence can help us to explore such things.
• self-management
  • I’m not exactly sure where this note could fit so for now have put it under self-management, but it may be more/also suited for e.g. collective-intelligence.
  • Anyway, as I was discussing the screen recording thing above and the personalised input generation stuff, I realised that what is probably an important thing to consider/configure is singular focus in any given domain. For example, in my personal mindware pursuits, there are many avenues of development I want to explore, but it’s easy to become overwhelmed with these options (paradox of choice) and not progress, due to all mental momentum being dispersed between all these potentialities.
  • So, I want to select a singular item from this list of desired developments to put my mind to, and persevere with that until it actually exists.
  • But more generally, it’s good to decide on one goal per life domain/project, and abstract away everything else as “eventually”/“to do later” so you can clear mental clutter and have more thought space for pursuing the current thing you’ve decided to complete.

Fri 01-05-2026

• mindware
  • A key bottleneck/issue I’m noticing with the automated transcription pipeline thing I’ve set up is that the transcription takes ages (for the length of audio I’m providing) and since I’ve set the task to only run when my laptop is plugged in (AC power source ?) then I need to be mindful that I don’t e.g. unplug or shut down while a transcription is running, or it’s just lost and has to be restarted (which is obvs a waste).
  • So, it might be good to e.g. manually set one going (when I know my laptop will be plugged in for a good while, uninterrupted) rather than have the transcription automatically triggered whenever an m4a file appears.
  • Alternatively, I can use e.g. my old laptop and/or desktop as dedicated “transcription engines” which run basically continuously, throughout the day (yeesh for the power consumption though).
  • Also, I want to trial using the small.en instead of medium.en whisper model, as maybe that could basically solve this issue (but it would need to have a similar accuracy to the medium model).
  • Anyway, this is definitely an issue I need to think more about, but also, I don’t want to get too tangled in these technicalities at the expense of considering the other functionality I want to develop (such as actually using and synthesising the transcripts), so I need to think about the most effective way to approach the development of these features (and handle various dependencies/bottlenecks).
• intrathinking
  • “intrathinking” is just a term TM currently uses for distinguishing independent/internal thinking processes from interdependent/external thinking processes (i.e. communication, thinking between minds), which is aka interthinking.
  • I.e., it’s to differentiate between thinking within minds versus thinking between minds (but obvs, both types of thinking majorly overlap and are inseparable).
  • Directed vs undirected thinking:
  • Anyway, something in the realm of increasing the effectiveness of my own thinking (i.e. intrathinking) I wanted to mention/explore is an approach of separating directed from undirected thinking acc directed-undirected
    • Often, I want to just let my mind “run wild” and follow whatever chains/trains of thought / tangents it feels inclined to - this is directed thinking.
    • But sometimes, I want to really drill into / hone in on a particular object (well, thought pertaining to a perceived/conceived object) and in this case, I want to minimise “mental distractions” and maintain a focus on this object, not allowing much if any diversion/digression (unless directly relevant to the current object of interest). This is undirected thinking.
    • Here, acc is just indicating this is “as currently conceived” as in, this is currently the best way I’ve thought of to represent this thought/idea/thing/concept/object.
  • So, what I want to trial is the initialisation of a “directed thought session”.
    • For example, when commuting, lately, I’ve just been making audio recordings of whatever comes to mind, no particular focus (undirected thinking).
    • But what I want to try is, before starting the recording, setting an “object of interest” to contemplate and express thoughts about. Basically, giving my thinking a direction for that segment of time (directed thinking duh).
    • And then, I can kind of alternate (not with any rigidity, just based on vibe of whether I sense I could do with a period of directed/undirected thinking) between these approaches, of either setting a direction, or not.
    • It’s an experiment!
• idk
  • This note is for idk - stuff I don’t (yet) know how to categorise/explain and don’t know where it belongs in broader overview of my worldview.
  • I think this note/object will have a lot of overlap with the imagine note/object.
  • A thing 10:02 extra
    • Overwhelmed. Too much thought generation. Tapping into the collective via integration, but perhaps the receiver has been momentarily refined/tuned to a greater extent than usual. Then started listening to the Odesza as it’s the music that reminds me of home/origin node the most, and it’s what generally soothes TM most when it feels most divergent (and hence “least belonging” on this planet) - it could be perhaps considered some of “most integral” music TM has so far encountered here. Anyway, now time to channel the intel into the studies/stuff in this lecture.