meltyness

While not as strict, you can filter lints by their description and apply a policy to your crate https://rust-lang.github.io/rust-clippy/rust-1.84.0/index.ht...

On the contrary, there used to be "the spoils system." I'm just a sideline listener but I think that's an appropriate lens to look at this through rather than some severe negative, or dangerously harmful trend.

I think trying to bar such a system simply would make the same things covert.

Pessimistically, if you have a product you want to sell to the government, this stuff is all just "veto bingo" where someone could turn you down, or move development.

Optimistically, it exposes clear marketable opportunities in existing suites.

This is also cathartic.

Is it also a disaster that many math education programs don't hinge on automated proof assistants?

"we just don't know" is a non-answer.

If Tao's point is in predicting rapid change from a game theoretic perspective, perhaps it would behoove university executives to court industry for research dollars, instead of an ouroboros of invented problems and federal tax dollars.

Even state tax dollars are more likely to have a clear purpose.

What these guys are on about: https://arxiv.org/abs/2301.10743

This is a case of doubling down on an investment too fast. Getting better lighting (enabled by cheap reliable LEDs) improved sales.

In this case the enabling technology price/existence was apparently not well-estimated. And apparently they didn't have enough foresight to use a conformal coating.

Also an initial contract of 10 years without a test market? That would be the pinnacle of incompetence.

I see this and conclude the opposite that they were adhering to the principle. Basically AI writes the buggy code that is upsetting you.

Similar experience trying to use GenAIScript, btw, and peering inside the box the code and product is pretty well incomprehensible.

The U.S. firms let everyone skeptical go the second they had a marketable proof of concept, and replaced them with smart, optimistic, uncritical marketing people who no longer know how to push the cutting edge.

Maybe we don't need momentum right now and we can cut the engines.

Oh, you know how to develop novel systems for training and inference? Well, maybe you can find 4 people who also can do that by breathing through the H.R. drinking straw, and that's what you do now.

Search and filter have been the paradigm for a while, since the book index was invented.

What are the tools of thought that will help students now succeed and advance technology in 20 years?

Perhaps there is a market fit for a laggard GPT. Something useful but exhibits clear errors that a student could pick up on, and so recognize the risk of blindly accepting information like this. The same issues are associated with excess reliance on traditional publications, "the armchair scientist" and so on.

Maybe, trained sequentially on the curriculum like a student, or something personally orthogonal so that the bot excels where the student is known to fail, create gaps where the student is known to succeed. It seems straightforward to fine-tune in this way since the corpus is smaller. For each grade level and setting A/B/C across the 4 subjects, maybe some 300 fine-tuning?

This was one for the (people who need to electronically hit) books.

I agree, lots of the body of human knowledge is wrapped up in natural reactions, moving visuals, textures, smells, and sounds.

The current batch is trained on just text afaik.

It's not like the market isn't aware of this. This kind of motivates middleware more generally, so that the design and functionality of custom services can be paid due care.

Definitely will take the Pepsi challenge on 'go doc' with 'cargo doc'.

Go doc had a bunch of, puzzling things going on / barely working, Rust doc is pretty much as described in the book and reference.

Although in the C ecosystem, Doxygen is pretty nice, the docs there have to be 3d to account for the way C code bases can work.

This seems pertinent, NEWS > bash-5.0 to bash-5.1 > mm

Seems to suggest it's indeed a hashmap in implementation, but I can't be bothered to look any closer.

They've got a YouTube with pretty robust updates. Iirc, during the lock downs, he'd do live coding and stuff.

So fun is proportional to the extent to which the punch bowl has been spiked? Seriously asking.

It still works, you can click on something non-focusable like the background or hit Esc and still get the "first page" thing. Like a touch interface.

Funny that's already so intuitive for how touch interfaces works, I guess it would take maybe a week to change but forever to decide what specifically precisely should be done, so if it's not something epoch-defining, world-changing and comparable to search itself, then maybe just leave it, lest you make a mistake that half the customers likes.

I was wondering why Google hadn't replaced 'I'm Feeling Lucky' with something to do with LLMs, or just added an LLM-generate option. I came to the conclusion that they're in corporate denialism over the whole thing. They'd be happier if their finger slipped and made Anthropic and OpenAI vanish until they could resurface and capture the market. Possibly not a great strategy.

It seems all of their years of letting the open web decay and vanish has caught up with the fact that many requests can be serviced with an inverse thesaurus manual snippet soup.

I wonder if there's a hyper-Turing test where an AI passes if the model, itself, cannot determine if it is talking to itself; or perhaps stated differently, maximizing some measure of control and processing duration to successfully conceal its identity under forced processing, discounting a solution that specifically learns to be silent or incoherent. I'm not sure what the value would be, just a passing thought.

This is probably already happening within the parade of censorship systems trying to imbue the models with agency

Again, probably pretty naive take (and without regard to the amount of or quality of outcomes), but when looking at the way academic research occurs, thinkers are often exposed to failures and asked to explain them or expound upon the underlying philosophy -- rather than to participate in the full engineering cycle; that can take different forms such as hands-on development, but particularly in engineering, viewed with skepticism.

I think there's even a take that you can tease the patent system as a way to profit from failing to explain ideas effectively.

Edit: maybe the consideration then is what are the roles missing? if there's a way to improve off-the-shelf model performance faster than moore, what is it? scaling to teams that specify more specialized roles; simulation, model architecture / quantization specialist, systems-level, hardware-match specialist / minimization? some sort of way to compose, or perform operations on the content of models?

Well, naively, then, what you might be describing is a new language; because if the software itself can't capture the details of the expertise, then that may imply that something simply isn't being expressed.

Languages evolve with expertise, specialties are tailored to and can be permanently and effectively captured; and not just captured, but reused as in the case of libraries.

Perhaps there's some core capability missing from the MATLAB ecosystem that isn't so obvious on the surface. What I know about MATLAB is it is primarily focused on simplifying and making performant DAQ / processing cycles, but not necessarily associating concepts to use-cases and correctness/effectiveness (and probably optimality, as an important thought) -- say with how Objects and Types compartmentalize data transformation effectively for user-facing software applications with understandable UIs.

There's kind of a product space to support "buyers", specifically calling out your section regarding "software integrator" vs "knowledge worker / domain expert." BTW "buy vs build" is a common engineering management language for this sort of thing I think; https://en.wikipedia.org/wiki/Component_business_model ... there's some silliness here in this space too since software itself is intended to already be maximally flexible; so "sellers" spending time building lots of flexibility is sort of a smell.

To motivate some of the machinery, you know systemically it's kind of a no-brainer for how financing and oversight works I suppose; overseers want to see early results which you might get from a knowledge worker rehashing or tweaking an existing solution, and then integrators get to play 'pick up sticks'.

So you have the integrator is the "buyer" because they have opted to defer product expertise (disregarding whether there's even any real cost, i.e., a FOSS solution, etc. "buy" meaning "not build").

The buy side has popular tools like linters, static analyzers; this is pretty huge. If you look at the "cloud native" space, treating lots of the cloud software players as "integrators" of software that's intended to be distributed, public, always-on, and at a cost of 'ad-supported' the buy side has the "Security and Compliance" layer: https://landscape.cncf.io/ ... a lot of money is flowing there too, since it simplifies precisely what you mention there, so you're in good company. Granted, "integrator work" might be a bit less specialized than the sorts analyses these tools might perform, but it's the same problem applied instead of abstractly "domain expertise" to "domain expertise of deploying existing software solutions."

It might not be popular thinking, but doc tools like CodeViz, doxygen, any worthwhile IDE, and other tools in the space are probably somewhere between the next space:

OTOH managers are sort of 'friendly buyers' for the expertise of their employees, so you can look at the tools of requirements elicitation, definition, capture, scheduling, prioritizing, and lifecycle management too, thinking specifically of like stories, epics, kanban, and presumably legacy systems like DOORS, which I don't have any experience with. If you want to avoid too much philosophizing and want to defer to some pretty broad experience SWEBOK has a lot of words about this more 'pre-compiled' approach rather than 'just-in-time' approach that most agile-type firms will use for their usually simple value props.

The same sorts of things might apply though; identifying whether "coverage exists" for "domain-expert produced prototypes" -- specifying the form of those, in say what you mention, signal processing. Basically, TDD -- does the shipped code match the tests? Does the integrator know enough to write such tests?

Essentially using 3rd party machinery to establish acceptance criteria, and clear targets to ensure that the "buyer-builder" contract is fulfilled.

Anything open-source can't really itself be a product, so you have to look at it from an incentives perspective. Companies open-source stuff like React so that the pool of capable developers for an internally-developed but, publishable tool can keep up with turnover, presumably without the expense of 6 months of onboarding.

I think that either condition, though, is probably monetization; either an upsell to more layers, or a downsell to the O'Reilly book...

Having recently gone through this trying to get into React ecosystem, but avoiding the most common frameworks, `create-react-app` has docs that sort of seem like a tutorial but rapidly spirals into random advertisements, since it's pretty simple and not really a React tutorial; in-fact Meta apparently disavowed the whole thing, but it still seems to work OK.

It seems natural, incentives and open-source are tricky though, since if you're trying to simply compete rather than collaborate, it will be counterintuitive.

In fact (you mention Next.js) React itself almost immediately spirals off into advertisements for 'supporting technologies' like some of the popular React-focused frameworks.

Maybe you can read either condition as, "Oh; this part doesn't do that" or "Oh; maybe I need to read the source code to understand" or an invitation to collaborate, rather than as a design flaw.

Hey this is brilliant thanks for the tip

Somewhere between tv and an infinite filing cabinet for bureaucracies; not to mention the holy grail of false productivity for the NYSE-conscious, an 'activity treadmill' for rosters of highly-credentialed people to seem concerned about.

Hey there, I'm strong with Networks and Linux. I might be a good fit to help optimize your builds, improve observability, or design and develop bottom-line features; perhaps you're building or expanding a Network-related app? Reach out, let's talk.

  Location: TX
  Remote: Sure
  Willing to relocate: Sure
  Technologies: IP Networks / Networking, Rust, Linux / Bash / Perl, Regex, Elasticsearch, Logstash, JavaScript, Ruby, Docker, Git
  Resume: https://github.com/meltyness https://blog.tomtech.link https://leetcode.com/u/meltyness
  Email: tjh@duck.com

Whose quarterly goal is quarterly goals?

TCP congestion control is a disaster. Setting DCTCP slightly improves things but for big slow pipes it can be impossible to get near line rate. Congestion control's intent also differs from QoS/CoS. It's also a 'reactive' design rather than a constructive one, usually a bad sign. Not to mention that TCP is itself really complex and susceptible to receiver resource exhaustion.

Still getting into the reading here, but making my intuition a little bit more explicit, it seems that at least the design, but possibly the implementation of Homa actually hinges on datacenter switches QoS queues being configured with a policy to explicitly respect the receiver-assigned priorities... cool!

Without regard to the detailed background, design or analysis, the approach of plucking a concern out of the Network/Data Link layer (prioritization / QoS) and moving it up to the Transport is a remarkably simple / clever start.

At least that's how I picture the start of the study of such a design.