azath92

I think there is an interesting reflection here in that software itself is pretty general, if we talk about a tool that the whole industry can use, then actually a general intelligence is to software engineers as the loom is to weavers is maybe actually an apt comparison?

Im speaking specifically of pre AGI general intelligence, little g general if you will.

I will echo what others have said and say this is amazing on a personal level, but also super cool to share with others. I think especially in tech we can get a little focused on efficacy and scale instead of human impact, and some might quibble about that in the comments ill be interested to see the discourse.

Your impact is undeniable, and i think quite importantly very tangible and concrete to you. Very inspiring.

Echoing the sentiment about this podcast in general. Its fantastic for well researched, tightly focused deep dives on some fascinating topics in fashion and fashion history. There are a couple of seasons that are long form one topic over the whole season, and a couple that are self contained episodic. Both formats are excellent, but for obvious reasons the self contained ones are a great way to pick one that grabs your fancy and see if its for you!

While this may be true, their throughput is practically unbounded. this has been a primary challenge to using local models at work, where its normal for me to have 1-6 sessions churning away at once. So while 100t/s on a single session is very fast, theres no way you could get a cumulative ~500t/s from a consumer setup like this, do the bandwitdth bottleneck on a single card/cpu/ram/ssd (tends to be gpu ram bandwidth is the limiter in my setup, but they all have respective limiters to this kind of throughput)

I think there needs to be some nuance to this.

Without the training data and open sourcing the training code and protocol you cant reproduce the model from scratch, thats true, but because the weights and structure are free to download you can modify, inspect and make derivative works off it. Calling it an inscrutable binary blob is misleading.

I do agree that its not reproducible and its often talked about like it is.

In some ways its like most scientific research with a "methodology section", which in todays academic climate often falls quite a bit short of a detailed step-by-step to reproduce results with all the ingredients, but in principle could be used by other experts to do this again. Even without publishing the exact data, a catalogue or description of what data and how it was cleaned/processed would go a long way towards a reproducible methodology i agree.

Ive gotten a lot of value lately using his writing to understand how parameter size, MOE and PLE variations on local models influence performance on various hardware. Its still a useful level of detail as someone who used to roll their own models for various applications pre-LLM but who now mostly uses LLMs rather than training them.

If you used to write pytorch, i think his writing is a great level of technicality to stay on top of things. And id look to them if i ever wanted to do a bit of learning/train my own.

Id love to hear if others have resources at a similar level of technicality and detial.

Id love an ELI5 for PLE. Im trying to work it into my back of the napikin math for compute vs memory bandwidth limitations on tok/s in PP vs TG work.

My attempt at a simplification of this article on it https://sebastianraschka.com/llm-architecture-gallery/per-la... into a couple of sentences is that they are linear embeddings of the input token space projected per layer, which are then gated by the transformer outputs per layer.

This would mean that the only one set of weights for the ple path needs to be pumped across the memory bandwidth as they are the same linear weights for all layers?

Sheit, maybe im trying to simplify something that i need to look at in detail. but id love to leverage others understanding if possible

Unfortunately, in my experience the more advanced/nuanced/complete of an explanation of physical phenomenon you want, the more math heavy and abstract the "real answer" becomes. Quantum Electrodynamics is some beefy stuff, the math is gnarly, and it is built on an increasing pyramid of mathematical reasoning and intuition as well as the steps of physical reasoning and intuition that you mention.

Remembering my graduate physics days im astounded that i was able to understand it, and am unable to really follow it now, as that higher math has gotten preeeety rusty.

Feynman diagrams are the attempt to loop back around and represent this field without all the math, but my experience was you have to have done the math first otherwise they are just so much arcane heiroglyphs. the wikipedia illustrates this well, where you can see a relatively clear diagram and corresponding equation, but i recall whole pages of a notebook to get one _line_ of an equation down. https://en.wikipedia.org/wiki/Quantum_electrodynamics#Feynma...

I say all of this not to suggest a clear explanation is impossible, but it sure wouldn't be easy. Love to see it though.

Some of the effort that folks have put into recording their dwarf fortress experiences, and captureing that special something that emerges out of the game is wonderful.

I humbly also submit the illustrated stories of Bronzemurder and Oilfurnace https://www.olddog.games/Oilfurnace https://www.olddog.games/Bronzemurder

and the story of Glasedcoast https://glazedcoast.neocities.org/

Love to hear of more. Even if im not actively playing, i always love a read.

If the goal is to provide a distinction between model and agent, i think the "agent system" is doing too much heavy lifting in the example here.

A useful extension to this mental framework that i use when trying to make this distinction is the application (cursor) -> which sometimes includes an orchestrator and all of the QOL stuff like resuming, checkpointing, etc. single or multiple agents (cursor agents)-> and runs a single or many agent instances (single agent in cursor)-> service api-> model.

This is to address a confusion i often see with agent being conflated with the application that we use agents in, rather than the distinction in the article which tries to unpick agent-model confusion.

For those of us on the other side of the ditch: all museums in London are free. and I would echo the sentiment that it makes exploration of them less of a completionist activity to "get your monies worth" and more of an enjoyable forey into a rich space.

Its literally a paradigm shift.

Also almost all commercial art galleries are like this, obvs for other reasons. The gallery district (google art galleries then wander those streets, chelsea is the most established) in NYC is an amazing place to spend a day wandering around seeing amazing art. my uncle who teaches art introduced me to it and i try to spread the word.

Free art for any reason is amazing.

I agree some why could be interesting. I would add to your list, because IMO this is more of a thought experiment-as-repo than a utility focused tool, what can i see by using this that i cant with a more fully fledged starting point?

I have to say I enjoy the brevity and clarity of this idea, given the ease of producing something large and unfocused in the last couple of years.

I hope it stays very simple, and would be interested if the author or anyone else can speak to what kind of complexity arises from using it, rather than bloating out the seed itself with initial complexity.

Ive been thinking about starting stimple with a pi framework for example and letting the tools emerge, but this is so much more lean to start, i wonder if it will be more or less interesting and or useful to start here.

Its also a common misconception that all glutamates are MSG (thats one salt of glutamate) and that all umami compounds are simple glutamate salts.

So while MSG is an excellent "neutral" umami ingredient, fermented products like this have a whole host of complex flavors that are often very desirable beyond just their glutamate salt.

Looking at the statistics that are used to measure growth and performance at a population level are such a great window into how they themselves are opinionated stances.

The two measures are both great measures of house ownership that measure different things. The article frames the second one as better, and if you are looking to see the ownership proportion of people (for example as they note to look at the numbers of young people who live with parents) thats great. If however you care about the housing development decisions that will incentivise owner occupancy you might care more about the standard owner occupancy per household.

I often see this with respect to GDP/GDP per capita vs mean income and other economic measures, but its great to see a non economic measure like this compared in this way.

That sounds like a tough realization to come to, and has been an opposite of my experiences in startups (as a non founder). Ive sought them out because the leadership cares so freaking much about what they are doing, and really genuinely wants to succeed. That can be intense, but at least you know they care. If you are early enough to have both equity and a solid chance at impacting the possibility of success you can have both mercenary and personal incentives aligned in a way that is impossible in a bigger company IMO.

I hope you give startups another go if that sounds good to you, and as you say, it sounds like you will be able to see the things you dont want at least. Good luck.

I can appreciate this sentiment. I am someone who favors a resign vote over a let it play in 8v8, and when i was thinking about why i think its that in a lobby based system like bar, where there are few games I value staying in a lobby rather than skipping around. Thus when a game feels lost to me, id rather play a new game (even to another loss) with a good lobby then play a slow inevitable losing struggle, or worse yet watch someone else play a losing struggle (even if they enjoy playing that loss out).

Something im trying to lean into in my games is to ask the players who are voting no what their win condition or what they think they can make a go of is.

I think what ive said above could make me sound like one of the folks who "want the tightest possible loop to winning and restarting", so to add some nuance there, I think playing a game where the conclusion is foregone lacks agency, and in a game where one team eventually loses of course there is a blurry area where the team agrees that they are in fact never going to win, and then another question on whether they lie to play regardless. Because of the variation there, i like the vote, and i am in favor of coms around why someone wants to resign or stay.

I actually think it is up to players in an open loby system to communicate their intentions and reasons, and while it is true that it means people who spam coms trying to push folks around are possible, i would echo ptaqs sentiment to use that report function! and on a softer level, i really like the mute. if someone is too much, mute them and continue to enjoy your game.

While i appreciate that this is a valid link. I would argue that its political stance is clear: its framing of this as "Biden’s border crisis", and its sourcing is poor: referencing fox news but no other news sources, and no links to primary documents or data, especially around the causality.

Additionally it doesn't appear to reference the actual HUD report from 2025 https://www.huduser.gov/portal//portal/sites/default/files/p...

This at least has AFAICT the 100% figure, but the causal link seems unsubstantiated. The exact, unreferenced quote from the document is "Immigration accounts for up to 100 percent of housing demand growth in some regions, and for two-thirds of rental demand growth nationwide. In California and New York, immigrants have accounted for 100 percent of all rental growth and over one-half of all growth in owner-occupied housing in recent years." p45 with no further citation to back that up. I know its not an academic paper, but some citation would be good to back this up.

Id be interested in more information on causal data, or at least a more nuanced stance on such a strong statement. Or a less politically charged comment.

Full disclosure, i (obviously from my comment) dont lean right, but one thing i appreciate about HN is a more nuanced comment, and i think this is a chance for that.

In a small team, or an aware team, where AI is being used all the time and we are figuring out the best way to do it, i often just preface my messages with

  - "from my ai to yours" where ive pointed my ai at some relevant context, and asked it to transform it for other ai context that a coworker needs
  - "my thoughts prettied by AI" where i just polished up my own words, often for outside coms, but indicating that i wrote the bones of it.
  - "i wrote this myself" in my case i tend to be very casual with my written coms, and ive been leaning into this in the past year rather than looking to correct it, as it gives the personal feel. but for cases where ive written more thoughtfully, i just flat out say that.

Now im not doing this rigerously, or obsessively, but i am finding it helps with exactly the kind of friction and erosion of trust that comes from reading things by ai as if i should treat it the same as a person and writing things as a person just to have it consumed and spat out again by an ai.

Helps my team is small. interested in how this could be translated to more widespread "company culture"

Id disagree with this analogy: "No carpenter is a specialist in drills." and i think its an interesting lens through which to look at the evolution of our tools.

I think there are trades where tool (or process if i may be allowed to extend the analogy) specialists exist and are highly valued. My dad is a plumber, so ill use that example but id trust similar is true for carpentry. there are specialists by task/output (new construction, repairs, boilers etc) but also tool specialist plumbers and companies for example drain clearing equipment or certain kinds of pipe for handling chemicals other than water are very specialised, and there are roles for them because the thing they enable, and the criticality of the task, and often the cost and complexity of using the tool are high enough to make specialisation valuable.

IMO software has, for the 10 years ive been working in it, been in an unusual position where the tools (languages, engineering practices, tech stacks) were super technical and involved, but also could be applied to a large number of problems. That is the perfect recipe for tool specialists: complex tool with high value and broad domain/problem space applicability.

Because of that tool specialisation, we've separated the application of the tool to a problem/domain from the tool use. reduction of complexity of applying these tools to many problems, means all domain specialists will use them, relying less on tool specialists.

imaging a mcguffin tool for attaching any two materials together, but which took a degree to figure out (loose hyperbole here), that sudenly you could use for 5 bucks and a quick glance at the first page of the manual. An industry that used to have lots of mcguffin engineers, would be mega disrupted, and you could argue that those tool specialists would have to identify more with what they were building than the mcguffin they were using.

the python cookbook is good. and fluent python is more from principles rather than application (obvs both python specific). I also like philosophy of software design. tiny little book that uses simple example (class that makes a text editor) to talk about complexity, not actually about making a text editor at all.

This is just the best. A very serious company, doing seriously cool and important stuff, also has an anime name/icon.

I wish more corps took themselves so lightly, while remaining serious about what they do.

I made the same gut assumption, and it points to either poor writing, or deliberately misreading writing that they mix units like that in the same paragraph, where presumably the idea is that we get a feel for growth in both?

Its probably nitpick correct, because the 12GW is planned capacity, while the solar might be measured use? but simple assumptins or conversions, as another comment points out, get you comparable numbers. taking the title into account, the whole article is a little bit smoke and mirrors on clear communication, despite having plenty of numbers. Thats a shame because it sounds like even unvarnished its good results!

Id guess by your smile there is an element of humor in your response, so this isn't a rebuttal, but rather i identified a lot with your point, and I was thinking that this is such a human response to vulnerability.

If it was guaranteed that it will not be abused or that I would regret it, it would not _be_ vulnerable. Just like its not bravery if I am not afraid or I am assured of my safety. Such a paradox. Being vulnerable for me is acknowledging that it might have an increased probability of a more negative outcome, but still trying to be vulnerable because of the huge connection unlocks that (often) occur in my experience.

On balance intellectually i am coming to see the expected value from being vulnerable in communications is high, but my little lizard brain keeps saying to me "what if you get hurt though" and being closed off haha. its an exercise to shut it up.

I find the money stuff newsletter by Matt Levine (bloomberg) great for this, the link is behind a paywal, but the newsletter is free. strong rec. todays newseltter https://www.bloomberg.com/opinion/newsletters/2026-03-11/pri...

From that newseltter:

> At the Financial Times, Jill Shah and Eric Platt report:

>JPMorgan Chase ... informed private credit lenders that it had marked down the value of certain loans in their portfolios, which serve as the collateral the funds use to borrow from the bank, according to people familiar with the matter. >...

>The loans that have been devalued are to software companies, which are seen as particularly vulnerable to the onset of AI. ...

From what i can tell the problem isn't that an individual who had cash to invest in a private (tech in this case) company goes down

the problem is that a company "private credit firms run retail-focused funds (“business development companies” or BDCs)" which took out a bunch of loans to invest in private tech companies is now having the underlying assets that they got those loans against (long term investments in private tech companies) valued lower.

the link im missing is what happens when people who also invested in BDCs want their money back, where their actual money is locked up in long term investments made to private tech companies, and their ability to get loans is now valued lower. I think this is called a "run" where if someone starts pulling money out, and ultimately you cant, then its a race to get your money out before others do, which applies to both the individuals and the institutional loans.

Note: my quotes are from the bloomberg newsletter i mention, which helped me, not the OP article. And i am writing as much to clarify my own thinking as from a place of understanding. I welcome clarification.

thank you so much! i suffered with this, and now i never will again!

ok, even that "few thousand examples" heuristic is useful. the usecase would be to run this task over id say somewhere in the order of magnitude of 100k extractions in a run, batched not real time, and we'd be interested in (and already do) reruns regularly with minor tweaks to the extracted blob (1-10 simple fields, nothing complex).

My interest in fine tuning at all is based on an adjacent interest in self hosting small models, although i tested this on aws bedrock for ease of comparison, so my hope is that given we are self hosting, then fine tuning and hosting our tuned model shouldn't be terribly difficult, at least compared to managed finetuning solutions on cloud providers which im generally wary of. Happy for those assumptions to be challenged.

Only to prompt thought on this exact question, im interested in answers:

I just ran a benchmark against haiku of a very simple document classification task that at the moment we farm out to haiku in parallel. very naive same prompt system via same api AWS bedrock, and can see that the a few of the 4b models are pretty good match, and could be easily run locally or just for cheap via a hosted provider. The "how much data and how much improvement" is a question i dont have a good intuition for anymore. I dont even have an order of magnitude guess on those two axis.

Heres raw numbers to spark discussion:

| Model | DocType% | Year% | Subject% | In $/MTok |

|---------------|----------|-------|----------|-----------|

| llama-70b -----| 83 | 98 | 96 | $0.72 |

| gpt-oss-20b --| 83 | 97 | 92 | $0.07 |

| ministral-14b -| 84 | 100 | 90 | $0.20 |

| gemma-4b ----| 75 | 93 | 91 | $0.04 |

| glm-flash-30b -| 83 | 93 | 90 | $0.07 |

| llama-1b ------| 47 | 90 | 58 | $0.10 |

percents are doc type (categorical), year, and subject name match against haiku. just uses the first 4 pages.

in the old world where these were my own in house models, id be interested in seeing if i could uplift those nubmers with traingin, but i haven't done that with the new LLMs in a while. keen to get even a finger to the air if possible.

Can easily generate tens of thousands of examples.

Might try myself, but always keen for an opinion.

_edit for table formatting_

I was aware of the patent, and agree i think its overly narrow and you could get around it easily. I think the reason we haven't seen it or something like it in another game (or i haven't but someone pleeeease id love to hear systems like it), is less because its not useful, or maybe its not useful as a plug and play because the only reason it works is because of the super exhaustive care taken on tuning its parameters and giving it enough variety to make it interesting to play.

Kinda like the dialogue/story paths in something like hades, where IIRC they made a whole system to manage it, but the reality is that system only matters when the tree is suuuuuuuuper complex. or maybe it was disco elysium, or both ...

And to separate my thoughts from the info blob:

i think the culture war point is also super true of the game design industry, not just the consumers, where the already ultra competitive nature of the work means that the creatives and the industry as a whole have taken a veeeery strong stance against genai. Thats a reckon, and i dont know if its good or bad.

It does feel a little counter to the march of progress, but in a medium where high effort can be enjoyed by many, im personally cool with artisinal handmade games.