mncharity

Fwiw, one case where I've wanted rollback has been input fusion over interface devices with diverse latencies. You might have 10 ms for a keypress, 100 ms for optical tracking, and 1000 ms for speech. So given click+"the red one"(spoken), you might start running click+"the one in front"(default), and almost a second later rollback and rerun with "the red one". Or for real example, keypress event handling might branch on optical "pressed where on the keycap" and "by which finger", which won't become available for several frames.

First, thank you for the comment. It prompted fruitful reflection. I note LLMs as nifty for this.

For me, success means "robust structural intuition". Perhaps frame it as understanding that's robust to adversarial noise? To fuzzing testing? If you fuzz content, changing numbers, inserting negations and lies, how extreme before there's a "wait, that doesn't make sense"?

Merely quantitative isn't sufficient. An Ideal Gas Law chapter problem, with numbers for solid Argon - mindless plug-and-chug - is not this success. But a sense of reasonable values, yes. Contrast the first-tier med student, asked for red blood cell size, who failing to recall it as a factoid, is quantitatively lost, retreating to "really really small".

Similarly, "descriptive" can be deep structure and constraints of a domain, focused on building structural intuition, or at least trying for it, or an embrace of "stamp collecting" focused on regurgitation.

I nod to "foundations for subsequent study, with the intended outcome that students have a broad but shallow understanding of core basic ideas. Lots of biology makes intuitive, mechanistic, and visual sense, much like [...] introductory chemistry. [...] without the broader context the understanding would be relatively fragmented, lacking the big picture." But then contrast it with content presenting a not-broad and quite-shallow take, that pervasively fails to engage with the domain's core structure. And then, even on its own shallow terms, still fails outcome-wise: First-tier institution students, coming to intro genetics from intro bio, lacking even a firm grasp of central dogma? Stoichiometry students not even thinking of atoms as real physical objects? So I see "wonderful books" and think "wat?!? - how about profoundly and pervasively dysfunctionally unhelpful books?".

Perhaps at root, there might be different visions of what a "big picture map" best looks like??? Maybe picture a human surface map, vs a USGS topography and geology one. Do details clarify by exposing patterns, or obscure as clutter? Does underlying structure? Do year-to-year research insights provide opportunity and motivation for frequent rewrites, or is there relative stability and slow evolution? Are labels and vocabulary treated as foundational, or as relatively unimportant? If you haven't seen part of the map, how important is being able to sketch it in plausibly? If fragmented into puzzle pieces, how important is being able to fit them together? How important is seeing why things are the way they are?

Maybe the contrast between a slim tourist guidebook, versus walking in conversation with someone who deeply understands the history and society and structure of a city? Both are accessible experiences. Conversation that's numerate will be richer than non. But while the guidebook can provide a bit of orientation, it's not even trying to leave you insightful and deeply clued in.

Thanks again. I'd not thought of the fuzzing analogy before.

> Get any modern undergraduate Intro Biology textbook [...] These are fantastic books

Curious how perspectives vary. I would have said there's basically nothing available, textbooks being horribly wretched.

I don't know of anything which takes a "bottom up", rough quantitative, engineering first-principles intro to cell bio, let alone to biology. No whys and hows of building close to thermal noise energy levels. No focus on pervasive multi-scale cross-cutting strategies for localization and compartmentalization. No energy budgets, not feel for reasonable numbers, no... sigh. When you see a nifty foundational insight mentioned in passing in a research talk, it's a really good bet it won't be in textbooks any year soon. One of the causal threads leading up to TFA, the Harvard bionumbers database, was born out of someone's 'it's absurdly hard to find numbers'.

Chatting with a cell bio tome publisher years ago, about what absurdly implausible resources would be needed to do something transformatively better, the snark for "but it has 100 authors!" was "nifty - and how many for the second page?". Maybe now with AI we can start nibbling away at this faster.

A cautionary note: the jiggle can be misleading, making one think motion is fully represented. These are traditional 3D animations, with their profoundly misleading handling of motion, just with a "jiggle filter" added.

For illustration, consider the classic animation of a walking kinesin towing a vesicle. One could jiggle-ify it. But that won't convey that during every step, the vesicle has done a "balloon in a hurricane" exploration of every possible position it can reach while remaining tethered. Won't clarify that the very very misleading "I'm just a peaceful barge" vibe is entirely animation fantasy. Secondary content could have been added to defuse this negative educational impact, but the choice was made to optimize for, and I'm quoting, "pretty".

Jiggle-ification takes perhaps the biggest educational downside of these animations, and makes it even more misleading.

A micrograph.[1] Usual EM prep caveats: surface appearance is all artifact, ~10% shrinkage.

[1] https://commons.wikimedia.org/wiki/File:E._coli_Bacteria_(73...

> order of magnitude calculations of about the processes of the cell

There's a searchable database of bionumbers[1], and a draft version of "by the Numbers" officially online[2].

[1] https://bionumbers.hms.harvard.edu/search.aspx [2] https://www.dropbox.com/s/gvpleqtcv8scro4/cellBiologyByTheNu...

Interesting. Tnx.

In case anyone else wondered about using gptel to edit thinking (eg vis Qwen3.6's `preserve thinking`), [1] explains:

> In a multi-turn request, from the time you run `gptel-send`, everything the LLM sends is passed back to it [...during tool calls...] includes multiple reasoning blocks. [...But...] subsequent gptel-send calls read their input from the buffer contents (or active region, etc), so the reasoning blocks in the buffer will not [] be sent as "reasoning_content".

But in org mode, those are apparently `#+being_reasoning` blocks (`gptel-include-reasoning`?), so editable thought might be an easy addition?

A caution, fwiw, that any llms which respond with interleaved content and reasoning blocks, currently only work when not streaming, and fixing that is non-trivial.[also 1]

[1] https://github.com/karthink/gptel/issues/1282

The xkcd color survey[1] was 16 years ago now. With the data available, there were various follow-ups. Many, including xkcd's own "take it for fun here" link there in [1], are now 404. But the strata[2] and word cloud[3] are still up, and relevant here.

The data remains available. It was a TidyTuesday ("social data project") for 2025-07-08.[4] This response[5] looked at TFA. (This week's TT is ag tariffs.[9])

I just did a quick and sloppy search, so there's likely more out there. Curiously, I found AIMode and Bing/chat less immediately helpful than I'd have guessed.

Years ago I used the xkcd data for a prototype web interactive for kids, which shader filtered video to selected colors... but it was overlapping rather than a partition, and I'd never leave out cyan. Hmm, maybe a vibe coding target.

[1] https://blog.xkcd.com/2010/05/03/color-survey-results/ [2] https://www.datapointed.net/2010/06/xkcd-color-name-strata/ [3] https://luminoso.com/the-color-cloud-an-interactive-visualiz... [4] https://github.com/rfordatascience/tidytuesday/blob/main/dat... [5] https://jofrhwld.github.io/blog/posts/2025/07/2025-07-09_col... Off topic: [9] https://github.com/rfordatascience/tidytuesday/blob/main/dat...

Just to stir thought, I note the TiddlyWiki[1] community (wiki as a self-modifying single html file; 20+ years old) has of course been exploring AI tooling... though not necessarily as an agentic environment. There's a markdown plugin, and others to make the file executable, or into a self-serving web app. Git is more problematic. So hypothetically, one could have a single-file agentic wiki wandering around and self-editing.

[1] https://tiddlywiki.com/

> and we couldn't come up with even one realistic example

This worries me about how we approach teaching rough-quantitative reasoning. It's widely thought of great importance for expertise. But even when there seems outlier opportunity (a first-tier institution emphasizing teaching, a dedicated class, for non-freshman, in physics), the questions... "Calculate how long to cook a turkey for, without reading the label or googling?" Really? Getting to a clear payoff may require a broader scope, or a different emphasis, or a reimagining.

> isn't any meaningful use of this step until it's combined with others

A kindergarten science educator suggests their students have a human right to understand their physical world then, not a lifetime (for them) later. Intriguing to think what that might look like as a goal... at any level.

Science education research, like teachers, is focused on the possible. Existing students, with existing resources, and constraints, and objective metrics. There's little incentive to stack counterfactuals: what if my incoming students had been taught A successfully, and B and C successfully, then done together, they enable D, and oh wow. Not when we're struggling to manage even one with any consistency. Disaster-triage chain-of-care doesn't encourage funding of population heath research. Especially when the possibility space is so little explored, that we don't even have a vision of what an E and F might look like. A incentive bootstrap challenge.

Two analogies for thought:

There's a meme, "Why develop one's own expertise? It's a poor investment. When you need it, you can hire it." Does AI make us all trust-fund kids?

An intro-physics educator at a first-tier university, observes that their entering students, having attended such well-funded schools, with such highly-skilled teachers, presenting material so clearly... widely lack both the skills and inclination to wrestle with a body of knowledge to extract their own understanding. To the detriment of their early-college education. The sci-ed snark version is, raising all students to the level of these, would be both an unimaginably immense triumph, and an ongoing profound failing to teach well. Will AI give everyone material presented so clearly?

> More oral defense of arguments. [...] More [...] work in which students must explain not only what a result shows but what it does not.

More rough-quantitative reasoning? Fermi questions. Especially if done by collaborative iterative bounding "Who can suggest another soft/hard upper/lower bound? ... What do you think of that argument?"

In contrast to a plug-and-chug theme, illustrated by an ideal gas law problem in a popular textbook, which despite years of use, and qc passes for multiple editions, has numbers for solid Argon. Reality checking, a feel for reasonable values, a "Is this approximation plausible here?", being pervasively "not on the exam".

Burst-parallel non-frontier models can resemble "tokens were free". And there one might potentially augment not just conversations, but CoT - retroactively by submitting messages with altered reasoning strings, or inline with the inference loop watching CoT and attempting non-distracting injection.

There's a much more active current thread[1].

[1] Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model https://news.ycombinator.com/item?id=47863217

Just brainstorming, but perhaps a more tangible gradient, with social backpressure? Imagine three identical patch tools: "patch", "submit patch", and "send patch to chief architect and wait". With the "where each can be used" explained or even enforced. Having the contrast of less-aggressive options, might make it easier to encourage more aggressive refactoring elsewhere. Or pushing the impact further up the CoT, "patch'ing X requires an analysis field describing less invasive alternatives and their un/suitability; for Y, just do it, refactor aggressively".

> allows agents to request (and later on drop) 'views' on functions and other logical pieces of code [...] The problem is that the models are not trained for this

Fwiw, I was playing with an "outliner"-tool collapse/expand idiom, on synthetic literate-programming markdown files, with #ids on headers and blocks. Insufficient experience to suggest it works, but it wasn't obviously not working, and that with a non-frontier model and very little guidance. Other familiar related idioms include <details>/<summary>, hierarchical breadcrumbs, and plan9-ish synthetic filesystems `foo.c/f.{c,dataflow,etc}`. One open question was comfort with more complex visibility transformations or sets - "hide #bar; show 2 levels of headers-only under #hee; ...". Another was cleanup - recognition of "I no longer need this and that".

Thanks for the reply. I had hoped my response to be encouragement, rather than ad hominem, as you never know when that might make a difference for someone. Sorry that didn't work out.

I suggest one source of the comment's downvotes, was not the views expressed, by the style of argumentation used. Though priors which read in a different subculture as not merely incorrect but silly, I agree didn't help. I'll focus on the argumentation.

Though as an aside, one difficulty in communicating between bubbles, is recognizing that "everyone knows" can become "that's absurd" at bubble boundaries. This can be funny. When Hollywood folks were complaining (re copyright law), on the record, about congressmen not "staying bought", rather than admitting to bribery, it was perhaps more something vaguely like "Everyone I talk with about this topic knows we're right.[bubble] Obviously and completely right. He knew what's right, so we funded him. Since he knows what's right, and did something else, the only thing which makes sense, is he was bought off to do what's not right". To do a cartoonish characterization. Anyway...

Yesterday, exploring your previous comment, I removed the link, and fed it to... either Bing/chat or ggl AIMode, prefaced with something vaguely like "The person who wrote the following wishes to escape their thought bubble. Evaluate and make suggestions to help them. "...comment..."". Thinking it might help me reply with something useful. The result was long and interesting somewhat (so probably bing). Maybe one-shot LLM chats could be a tool towards a goal of bubble busting.

The item I recall now, was noting a "morality play" style of argument, with cartoon characters ("liberals" this, "scientists" that). Imagine in some profession context, you're looking for someone to provide reasoned technical analysis of some complex system, and you get "oh, it's straight forward - <large group> just needs to confess their sins". I won't here critique that as theology, but as analysis, what would think of that? What value would you place on it? Or in a discussion of some web phenomenon, a comment "it's simple - web devs are bozos - they just need to pay attention". There is a, lack of fruitfulness, in this flavor of abstraction. So I suggest "dissenting opinion on this liberal site" misses one likely source of the comment's downvotes.

Thanks for the conversation.

Recent related discussion/blog[1].

[1] The acyclic e-graph: Cranelift's mid-end optimizer https://news.ycombinator.com/item?id=47717192

Glancing at your comment history, I see you're already familiar with the concept of groupthink bubbles being extremely popular with people, and a big business. And you mention a wish to mitigate them. I'd hoped to write something more useful here, but it's very late, and drafts are just not gelling. So merely... I suggest the above comment provides opportunities for such mitigation.

I revisit the site's open tab, and find a full screen, black background:

> It is 12:56am. Do you still want to be on the internet? Yes No

Lol. Nice touch.

> No -> "Okay", fade to black.

Nod. Mine was VR dev leftovers. Fwiw, running 6ish prompts in parallel, roughly doubles my aggregate t/s (but requires cooling kludgery). If one's goal is not local, but rather real-time or consistent or transparent or scalable, there's AWS.

One sidebar caveat re "René Girard observed that human communities in crisis often resolve internal conflict through scapegoating", and "humans will scapegoat at all times", and treating this as "human liturgy", rather than as culture-specific pathology.

Consider a South American fishing village culture, which didn't do concepts of "shit happens", nor "unintended consequences", nor "systems failure". So everything that goes wrong, is the result of some one individual's intent, exercised by direct action or by supernatural whammy. Your neighbor. Your family. Your enemy. Some fish escape the group net? Gather around to figure out who did it. Your plant not growing well? Stub your toe? Pay the witch doctor to tell you who whammied it/you, and to whammy them back. Described as a singularly toxic interpersonal environment.

Now moving subcultures to a safety/learning/challenge culture is so very hard (aviation CRM, etc), perhaps it's a valid approximation to view its absence as fixed. But having just read a comment suggesting unaddressed climate change should be blamed on the science community because reasons, I'm sensitized just now to "our local cultural dysfunction is simply human nature - it's them that's the problem". Even acknowledging bad actors fueling the dysfunction in which they thrive.

> "I think one of the most aggravating, restricting facets of lunar surface exploration is the dust and its adherence to everything no matter what kind of material, whether it be skin, suit material, metal, no matter what it be and its restrictive, friction-like action to everything it gets on [...] the simple large-tolerance mechanical devices on the Rover began to show the effect of dust as the EVAs went on. By the middle or the end of the third EVA, simple things like bag locks and the lock which held the pallet on the Rover began not only to malfunction but to not function at all. They effectively froze. We tried to dust them and bang the dust off and clean them, and there was just no way. The effect of dust on mirrors, cameras, and checklists is phenomenal. You have to live with it but you're continually fighting the dust problem both outside and inside the spacecraft. Once you get inside the spacecraft, as much as you dust yourself, you start taking off the suits and you have dust on your hands and your face and you're walking in it. You can be as careful in cleaning up as you want to, but it just sort of inhabits every nook and cranny in the spacecraft and every pore in your skin [...]" Eugene Cernan, Apollo 17 debrief[1]

An interactive microscope of regolith.[2] Like tiny broken glass, hard as rock, and sticking to everything like static-charged packing peanuts.

An old tech memo and paper.[3][4]

[1] https://an.rsl.wustl.edu/apollo/data/A17/resources/a17-techd... page "27-28" 258, 50 in pdf. Lots of other mentions of dust. [2] interactive microscope of regolith https://virtualmicroscope.org/sites/default/files/html5Asset... [3] The Effects of Lunar Dust on EVA Systems During the Apollo Missions https://ntrs.nasa.gov/api/citations/20050160460/downloads/20... [4] IMPACT OF DUST ON LUNAR EXPLORATION https://adsabs.harvard.edu/pdf/2007ESASP.643..239S

Hmm. Perhaps there's a niche for a "The Missing Guide to llama.cpp"? Getting started, I did things like wrapping llama-cli in a pty... and only later noticing a --simple-io argument. I wonder if "living documents" are a thing yet, where LLMs keep an eye on repo and fora, and update a doc autonomously.

I hadn't tried that, thanks! I found simply defining GGML_CUDA_ENABLE_UNIFIED_MEMORY, whether 1, 0, or "", was a 10x hit to 2 t/s. Perhaps because the laptop's RAM is already so over-committed there. But with the much smaller 4B Qwen3.5-4B-Q8_0.gguf, it doubled performance from 20 to 40+ t/s! Tnx! (an old Quadro RTX 3000 rather than an iGPU)

Fwiw, with its predecessor's Qwen3.5-35B-A3B-Q6_K.gguf, on a laptop's 6 GB VRAM and 32 GB RAM, with default llama.cpp settings, I get 20 t/s generation.

Oh, awesome. On my doables list was to try combining text tokens with "scent" embeddings, to give LLMs a higher-dimensional reading experience. In a file listing, larger files might smell "heavy" or "large". Recently modified files "untried" or "freshly disturbed". Files with a history of bugs, "worrisome". Complex files might smell of "be cautious here - fragile". Smelly `ls`.

Or, you might save token sampling telemetry (perplexity, etc) alongside a CoT and result. So when read, it's like a captured performance - this sentence smells "hesitant", that one "confused". Poetry vs prose. Or, a consistentcy checker might add smells of "something's not right here". Or... emojis that emote.

For a dog, that's not merely a lamppost, it's richly-evocotive local history. To a dev long experienced with some codebase, that's not merely a filename, it's that nasty file that bites.

One open question is whether you can find and calibrate embeddings to provide an informative whiff, without badly degrading reasoning. And be cautious of, and suspicious of changes to, a scary file, without becoming too avoidant. Also, salience bias. Also, imagine debugging scent hallucinations.

Activation-rich text - auxiliary non-linguistic embeddings as meta-signals... the random silliness local LLMs encourage.

Hmm. HTML has always had goals and tradeoffs which are in tension with many uses. XML too. Witness the very many versions of "write this instead, and it becomes HTML" - long and widely used and valued. Perhaps we collectively might have done better, but we didn't. Turns out LLMs also find different formats significantly easier to use for different things.

As a tradeoff example, yesterday I again tripped on the KISS "CDATA doesn't support HEREDOC-like prefix whitespace removal". So does one indent, compromising payloads where leading ws is significant, or not, confusing humans and llms.

Re reinvention and first principles, aside from engineering tradeoffs, it can be hard to understand design spaces and to be aware of related work. I suspect there's a missing literature to support these, but professional organizations have been AWOL, and research funding dysfunctional. And commercial conflicts of interest. And it's hard. But now coding LLMs are messing with "don't reinvent wheels" payoff tables. Perhaps we'll someday be able to be explicit about design space structure and design choice consequences too. And perhaps we're already getting transformatively more flexible around format extension and interoperation. TFA isn't just a new format - it's a github repo which will help teach LLMs how to do progressive execution of fenced code blocks, making the next format which does this potentially easier to create. "Merge in what X does, but <change request>". Yay?

IIUC, non-meme carcinization is something vaguely like "similar tradeoffs pressure towards similar forms in diverse contexts". LLMs might help us more easily understand tradeoffs, implement forms, and manage diversity?

> AI chats as a valid document

So many formats, with different tradeoffs around readable/parsable/comments/etc. I wish there was a "universal" converter. With LLM's sometimes used to edit chat traces, I'd like ingestion from md/yaml, not merely a "render from message json".

So .json `[{"role": "user", "content": "Hi"}` <-> .md ` ```json\n[{"role": "user", "content": "Hi"}` <-> above ` ```user\nHi` <-> `# User\nHi` <-> ` ```chatML\n<|user|>\nHi` <-> .html rendered .md, but with elements like <think> and <file> escaped... etc.

> If you're still looking for a name let me suggest "hyper text".

Perhaps "WWW SPA document"? Using markdown with highly-progressive fenced blocks?

Hypertext (one word, coined 1960s) is quite a broad category. Subcategory "WWW" could fit, as TFA seems WWW-ish. A markdown document format, and progressive rendering of tags and code, seems HTML-like. Though with greater progressiveness - code blocks with streamed execution rather than merely compilation. The progressive JSON callbacks, React, integrated client and server code execution, and server-side rendering, seem closer to WWW SPA than to HTML. Though SPA files often seem more "source" than "document". And the multiple-page "App"-ness of SPA doesn't fit well. SPA seems a better fit than "full-stack". Perhaps some name analogous to "isomorphic javascript"...?