jvanderbot

Yeah, no notes. It's remarkable.

TL;DR: Wired up an in-game predictor of RNG output and used it to only craft legendary items when RNG would line up to roll legendary.

From base to legendary at a suprisingly high rate. Look very closely at the video at the top of the post - what I was seeing didn't sink in until I had finished the article. Amazing.

Well given it took me 1 minute to type that, I consider it a loss leader.

This is 100% a valid workflow in DoD/ITAR fields. Suffer once through review, benefit eternally. It's especially important to get one good reviewed spec / presentation out, then all later reviews can be staged w.r.t. that one and are much easier to baseline and pass.

That'll be $1000 consulting fee on ITAR compliance.

DDS is hard. But we have agents now.

It does, thus (1)

You're right. In rust, IMHO, the problem is mostly solved by using system C libraries rather than trying to force Rust into something that looks like the optimized C/Fortran we've had for decades. In almost all cases, it's quite possible and quite painless to do it that way. For example, by just making it a matrix operation and letting the system wrapper library call out to something blas-like.

Very generally speaking, you want to view dissasembled sections.

You can do this at the function level.

In rust, try this: https://github.com/pacak/cargo-show-asm

However, life is much easier now, because an agent knows how to steer through this with you. Have it teach you at first, then you know how to ask.

This is one of those "back in my day we had to ... " stories, btw :D

I do this. I do this a lot. My job involves processing a significant amount of weather data and flight telemetry _very quicky_.

What I have found is that getting rust to auto-vectorize is a nightmare. The options available to me have always converged around: 1) use simd-like apis 2) frame it as matrix-vector operations.

(1) is touched on in TFA

(2) is way easier, and allows use of well-tested apis and crates, each of which (sensibly) call out to better-tested C libraries. Each of those can, should, might, or will use your CPU better than you will. If you can frame it as matrix-vector operations, you will go very fast, not least of which, by stacking the operations into a _big_ matrix/vector op, which your CPU will happily tear though.

However the article proposes a third, very cool option: use algebraic ops API! Worth a read.

   Rust used to not have any reasonable way to do anything like this on stable (for my own definition of reasonable), but now it does! Rust 1.98 stabilized algebraic operators for floats. These allow you to declare per-operation that you’re okay with the compiler making optimizations that may change the result as long as the optimized code is algebraically equivalent to the original. Yes, this includes potentially reordering operations.

I'm absolutely sure there is. Any poorly specified constraint will be abused.

It would seem there are! The question is: Will there continue to be?

Of course not. I am preparing for a 30% cut in 3 years, but don't necessarily expect it. It's not a guarantee that any demand at all will go down, or that AI will do anything whatsoever to our job. Demand is up lately for senior devs (and junior devs probably have more of an evaluation issue than closed job reqs[1]).

[1]: Sorry for the blog post: https://jodavaho.io/posts/ai-signalling.html

Yeah, fine. Honestly I think most of us have been overpaid our entire career. But I'm not sure I see it necessarily. Take these stats, which can cut either way on this topic:

In USA there are more software engineers than all other engineers combined.

There are 10x more software engineers than the next highest engineering discipline.

If you could subdivide s/w engs into 10 equal categories ("I'm not just a run of the mill s/w eng"), each of those would be the top engineering employment category.

The top 10% of software engineers outnumber the next highest engineer class.

At 10% unemployment, there are more unemployed software engineers than the next highest employed engineer category.

Both can be true.

Finding the top 1% gets better results if you have a wider net.

Finding the cheapest 20% does too...

Nothing to special, but I made a workout tracker that allows me to only log a standard, scored workout.

This solves exactly n=1 user's issues: I tend to fall off the workout wagon without public accountability and when I dont get scores to improve on and especially to degrade and dillute my gym trips to the point of ennui.

It's free and public until it breaks:

https://aft.jodavaho.io

My progress

https://aft.jodavaho.io/track/46B8XWPCVK5MVS35R8X53Y8KY5

(heavily helpd by AI on the frontend - I am terrible at it)

This is like a bizaro second amendment debate for the 21st century.

Exactly right.

When local models and startup labs can distill / learn / accelerate open models for local use by startups - the rational response by incumbents is to call LLMs doomsday machines that cannot be trusted in the hands of normies.

Everyone addresses this differently, but for me, I always grated at people whose knowledge of esoteric patterns and language features was their whole identity.

It feels to me like suffering through programming built that knowledge, and so that knowledge was an indicator of taste and experience and persistence, but then it became cosplayed as people sought the knowledge itself in "patterns", or "features" rather than the experience of just building things.

The LLM thing just cuts the same way. It's probably not using your favorite pattern or language feature, or maybe it automatically is, but either way your annoyance is about it stems from the quality signals being discounted: You can't really assert experience through that kind of knowledge anymore if people are providing value while stepping around your hard-won lessons and using your favorite pattern or feature through LLMs. You saw this in the linkedin posts early on: All the programming influencers asserted their value by saying they wrote better quality code than the code the LLM wrote.

This is worst for junior devs. Making a clean, well factored codebase as a portfolio project was a near-instant win on the job market because it meant this person either suffered like you did, or they listened to seniors and walked the walk. That signal is absolutely gone now.

I think that dilution of signal explains almost all of the jr hiring freeze, actually: https://jodavaho.io/posts/ai-signalling.html

Yep, that's how skill stacking works.

    Generalist > specialist.

No, my friend, if you are deeper than "just a pilot" and also capable of defining and building software (even with claude) that works and works well as extensions to industry tools like CRM, then you are not a generalist, you are an intersectional specialist of the kind Scott Adams used to talk about as "skill stacks" here:

https://www.youtube.com/watch?v=rG3PcdTyZ6w

You have two good skills, median-level among those specializations probably, and combine them into a top-tier niche ability.

Also, did you voluntarily disclose that you thought that comment was about you? It feels like you are the friend they were talking about, lol

Have subagent periodically review the work and plan.

Does anyone remember the FTC "comments" about net neutrality? And how completely made up it all was?

It must be that everything is like this nowadays.

You're gonna laugh, but...

I ask alexa and it generally does fine. If uncertain I go outside for a bit and look at clouds.

For the details, you just have to know how your local weather works - thunderstorms come in squall lines here, or are hot-day flash storms.

Blizzards follow warm still weather the same way, but mostly in Jan-Mar, when the jet stream curls away and will snap back to bring arctic air down.

Moisture (rain/slush potential) follows California rain by 2 days or so. If my friends in cali complain about rain, I know we're gonna have school closures if in the winter.

If I really need to drill in on rain/snow chance, I google for a weather radar and do prediction in my own head. (windy.com is ok, accuweather is real data to watch). Storms generally move linearly or rotate, and gentle rain is wide while torrential rain is clustered.

Locally, many of the old wives tales about the shape of clouds or direction of wind vs prevailing (with against, cross-grain) are true enough for 12h-24h predictions in upper midwest USA.

My work involves parsing real-time weather data at an existential level - because our entire use case is built around providing weather-related alerts and re-routes for aircraft.

The accuracy of weather data has been largely unaffected since DOGE took hold. The reliability of the infra, however, has suffered somewhat. There are more outages now than there ever were, with a massive spike in outages, not kidding, the month DOGE was at its peak. At that time we even saw significant manifest changes - stupid stuff like clearly LLM-generated json fields that were whole sentences instead of the key/value pairs from the schema. It was a wild time.

Anyway downstream we're mostly fine nowadays. I can still follow storm tracks accurately, see real time responses, correlate with pilot reports and lightning strikes, compare with aircraft telemetry mostly favorably, etc etc. It's quite remarkable what high-end weather products are able to do, the ones you have to work for.

We do, however, merge several free/paid sources into one picture. Rarely do I find serious disagreement, but it does happen.

Or worse, the comments are a tutorial through the changes for your review, referring to something that now doesn't exist because of the changes you're making. Less than informative, literally backwards facing.

But that's what stackoverflow looks like.

Oh, that's somewhat philosophical. The prediction here is a mechanical thing, just random sampling along a pre-calibrated mapping of options given the tokens and their importance weights. That sampling can be called prediction if you use it that way, or it can be called thinking if you are into the psychosis too deeply. Or it can be called a CLI tool command if you're in an agentic framework

In what sense is the model not predicting?

Yeah the chinese room "fallacy" fails to account for a chinese room that can do tool calls and network traversal.

But it is a next token predictor.

Recursively invoked.

With carefully selected context.

And massive investment in RL to tune token selection.

And the ability to use cli tools on other folks' machines.

That's a powerful system built around a conceptually simple technology: Next token predictors.

A better statement might be:

    Current agentic systems may be *built* from next-token predictors which are conceptually simple, but because of agentic frameworks, recursive invocation, tool use, and *heavy* investment in reinforcement learning in these contexts and for specific applications, they can no longer be thought of as "Merely" next token predictors.

Modern agentic work is probably more of a "emergent system from simple rules and complex interactions" paradigm than a genuinely new technology.

"cui bono"