llm_trw

The point is that you can force the respective of a giant or an ant if you so choose.

Using two cameras I had a time lapse of a fog coming in where you could see the 3d structure becuase they were at two different streets looking out.

Like op this was done nearly 20 years ago. Now you don't even get fog any more.

And people obscure this fact by talking about spending controlled by XYZ instead of total federal spending.

Who cares that it's an act of congress that funds a given project instead of a regular budget? Your money is still being spent.

Here is an infographic from 10 years ago that shows where the money is actually going: https://visual.ly/community/Infographics/politics/death-and-...

In short: if you're not auditing the DOD you're not trying to fix anything.

If I was sending it as a business email I'd edit it before sending it off. But the first draft saved me between 10 to 30 minutes of trying to get out of the headspace of "this is a fucking disaster and I need to look for a new job" to speaking corporatese that MBAs can understand.

This is where the HI comes into the equation.

>The second incorrect method to save a matrix of embeddings to disk is to save it as a Python pickle object [...] But it comes with two major caveats: pickled files are a massive security risk as they can execute arbitrary code, and the pickled file may not be guaranteed to be able to be opened on other machines or Python versions. It’s 2025, just stop pickling if you can.

Security: absolutely.

Portability: who cares? Frameworks move so quickly that unless you carry your whole dependency graph between machines you will not get bit compatible results with even minor version changes. It's a dirty secret that no one seems to want to fix or care about.

In short: everything is so fucked that pickle + conda is more than good enough for whatever project you want to serve to >10,000 users.

It is heaven.

With a well working rag system you can find the reason why any decision was made so long as it was documented at some point somewhere. The old share point drives with a billion unstructured word documents starting from the 1990s is now an asset.

No. I'd already spend 30 Min looking at how to solve it myself. The search on perplexity was a hail marry before I started calling handymen.

Linus Trovalds: bottom 20% coder.

https://www.youtube.com/watch?v=VHHT6W-N0ak&pp=ygUMTGludXMgQ...

I fixed the hinge in my oven by giving perplexity.com the make and problem. I saved an hour on the phone calling people to organise a visit some time in the next week.

Maybe you should stop using the Ai slop tools that don't work?

Counterpoint: the PC.

The minicomputer revolution was meant to replace all the big iron in the basement with computers shared by teams. This lasted a decade until we got computers meant for one person and we've used that format of computation for nearly 50 years now.

Everyone is talking about how AI will replace people.

It won't.

AI without HI is useless. However, HI with AI is much better at handling all the bullshit office work that needs to get done—anyone who doesn't understand this is either an idiot, a troglodyte, or already has an expendable army of minions doing their work.

The surge in AI demand won't be top-down but bottom-up. That stupid office birthday wish card being passed around? I can Shakespeare the motherfucker. And that email to the big boss telling him the project is doomed and that he needs a compass to find his own ass? Let's ask R1.

>My boss is a moron who can't find his ass without a compass. His latest tech initiative will cost us millions and make the best staff leave. Write a professional letter that covers those points without being harsh.

---

Subject: Request for Review of Upcoming [Tech Initiative]

Dear [Recipient Name/Team Name],

I hope this message finds you well. I wanted to share some thoughts regarding our upcoming [Tech Initiative Name], as I believe open dialogue will ensure its success while aligning with our team’s strengths and long-term goals.

While the initiative’s intent to innovate is commendable, I’ve observed several potential challenges that may impact its effectiveness:

Financial Impact: Based on preliminary analysis, the projected costs appear to be significantly higher than initially estimated. A detailed cost-benefit review or phased rollout might help mitigate financial risk while allowing us to test feasibility.

Team Retention Concerns: Many of our top performers have expressed reservations about the initiative’s current structure, particularly around [specific pain point, e.g., workflow disruption, lack of clarity]. Retaining their expertise will be critical to execution, and their insights could refine the plan to better address on-the-ground needs.

To ensure alignment with both our strategic objectives and team capacity, I respectfully suggest:

Conducting a collaborative risk assessment with department leads.

Piloting the initiative in a controlled environment to gather feedback.

Hosting a forum for staff to voice concerns/solutions pre-launch.

I’m confident that with adjustments, this project can achieve its goals while preserving morale and resources. Thank you for considering this perspective—I’m eager to support any steps toward a sustainable path forward.

Best regards,

Man if we only had some type of markdown meant for machines to understand, that was specifically designed to handle arbitrary information nesting and tagging our lives would be so much better now.

We could have called it something like extended markdown language or something and use a wicked acronym like eXMaLa for it.

Shame no such technology exists and never did.

The best argument against the EU is a 5 minute conversation with anyone who supports it.

That you (and apparently the German bureaucracy) think people who can start a business and people who can tick administrative boxes are the same type of person explains German economic performance in the last 5 years.

How long did it take you from start to finish?

In Australia, Canada and the US I can open a company in about 30 minutes from picking the name to opening a bank account I can get money into.

In Germany the last time I tried it was about 7 weeks.

There is a huge gap in the market, I'd say between capital investment of $50,000 and $5,000,000, where a ton of ideas which can make 20%+ ROI per year once mature live. But because they will never hyper scale they can't be financed by startups, and banks are too scared and stupid to invest.

I've done the math for a document segmentation pipeline that every RAG system needs and you'd be able to get one off the ground for around $2,000,000 USD and be rolling in cash instantly. Good speaker diarization is another family of models that will print money once solved, but again, there isn't the possibility for hyper scale.

Meanwhile at least a dozen companies I've spoken to have wasted on the order of $10,000,000 trying to solve these problems in house and failed.

Try opening a business in Germany to find out why.

Because is the bitter lesson were true no one would be wasting their time with convolutions or attention blocks. You'd just replace them with the general tensor that allows every hyper relation possible between all points instead.

Teenage me who wanted information to be free didn't imagine this would be the result.

It's like everything I wanted to happen in the 00s did, but from the monkeys paw.

The only way I can think of fixing this is by giving rights to flesh and blood people that corporations don't have.

GPL? Humans only.

Free speech? Humans only.

It was the only place where you could purchase the ebook.

I contacted a well known author about the shit latex rendering of his ebooks on Amazon. He said sorry and send me the a pdf copy he build with my name on all the pages. I really like the fact that I have a book dedicated to me by the author, but why the fuck do we need Amazon in this interaction?

The torrent sites is how meta and friends get their books. They are safe.

I know they do good stuff, it's just the fire hose of bullshit is a lot wider than the tap of innovation.

I don't know. Google has some great papers but a lot more hot air. The best thing to do when you see a google paper is ignore it until someone you trust more writes a second paper about it.

Well it was a nice run.

I expect all LLMs to be illegal within a year if this is the sort of high caliber defense the top minds of meta can come up with.

It's a tough one, I've never seen a book that actually covers the _bare_ minimum of the maths you need for ML.

The little learner comes close but I'd only really suggest that to people who already know the maths because the presentation is very non-standard and can get very misleading.

If you're interested drop me a line on my profile email and I'll have a look at some numerical algebra books and papers to see what's out there.

I mean yes, but without knowing the maths then knowing how to optimize the maths is a bit useless?

At the very least you should know enough linear algebra that you understand scalar, vector and matrix operations against each of the others. You don't need to be able to derive back prop from first principles, but you should know what happens when you multiply a matrix by a vector and apply a non-linear function to the result.

The only reason why suburbia is possible is because it is _heavily_ subsidized.

If you had to pay the real bill for road maintenance alone suburbs would no longer be viable.

So the suburbs take from the commons and don't give back in your example.

>Interesting, how are you remixing the order of questions? If we're talking about an academic benchmark like MMLU, the questions are independent of one another. Unless you're generating multiple answers in one go?

Short version: if a model can answer a very high proportion of questions from a benchmark accurately then the next step is to ask it two or more questions at a time. On some models the quality of answers varies dramatically with which is asked first.

>Docs for synthesizer's here: https://docs.confident-ai.com/docs/synthesizer-introduction, there's a nice "how does it work" section at the bottom explaining it more.

Very good start, but the statistics of the generated text matter a lot.

As an example on a dumb as bricks benchmark I've designed I can saturate the reasoning capabilities of all non-reasoning models just by varying the names of objects in the questions. A model that could get a normalized score of 14 with standard object strings could get a score as high as 18 with one letter strings standing for objects and as low as zero with arbitrary utf-8 character strings - which turns out mattered a lot since all the data was polluted with international text coming from the stock exchanges.

Feel free to drop me a line if you're interested in a more in depth conversation. LLMs are _ridiculously_ under tested for how many places they show up in.

>This brings us to our current limitations. Right now, DeepEval’s primary evaluation method is LLM-as-a-judge. We use techniques such as GEval and question-answer generation to improve reliability, but these methods can still be inconsistent. Even with high-quality datasets curated by domain experts, our evaluation metrics remain the biggest blocker to our goal.

Have you done any work on dynamic data generation?

I've found that even taking a public benchmark and remixing the order of questions had a deep impact on model performance - ranging from catastrophic for tiny models to problematic for larger models once you get past their effective internal working memory.

Which is my point.

You need a ton of specialized knowledge to use compute effectively.

If we had infinite memory and infinite compute we'd just throw every problem of length n to a tensor of size R^(n^n).

The issue is that we don't have enough memory in the world to store that tensor for something as trivial as mnist (and won't until the 2100s). And as you can imagine the exponentiated exponential grows a bit faster than the exponential so we never will.

I am in correspondence with a number of worth while authors, it's just that there isn't any place where they congregate in the (semi) open and without the weirdos who do stuff with the models you're missing out on a lot.

My favorite example I can never share in polite company is that the (still sota) best image segmentation algorithm I ever saw was done by a guy labeling parts of the vagina for his stable diffusion fine tune pipeline. I used what he'd done as the basis for a (also sota 2 years later) document segmentation model.

Found him on a subreddit about stable diffusion that's now completely overrun by shitesters and he's been banned (of course).