dada216

49 million people live in Spain. and it's a non industrial economy.

they will raise taxes while cutting the budget. in a multiparty system there are ways to achieve politically unpalatable things. in Italy we have what we call "technical governments", we don't vote for the government, only for parliament, and they elect the government, when nobody has a clear majority they vote for a "compromise" government where they elect a prime minister that is from no party, usually an economist or someone like that, which carries out the politically unpalatable reforms, in France they will do it with a split parliament/presidency

republican government. I wouldn't call that a democracy, not in the modern sense of the term.

that 200 years average does not sound correct. and so the rest actually.

You can't really predict how a collision would happen and hence where you should land.

I'd say the best way to avoid collisions at scale is already on the road and its cars with sensors and computers, if we weren't living in a pre-cyberpunk dystopia I would gladly have them coordinate to achieve that

I would say quality of the training data is probably the most important one. LLM write better C than Rust in my view, and I think it's because the quality of the training data.

I wonder if even people like Intel or AMD or Nvidia do.

There might be a strict hardware audit process for cryptographic accelerators on IBM Z machines though, I'm not sure, I'm gonna ask though.

The snark is mostly driven from distrust in Anthropic.

I see. We are picturing a malicious LLM here that Im basically trusting. But zero trust is a thing.

Those can still be handled by system tooling, setattr in the .git dir that prevents executable permission, container binded to a network on a VLAN that is in DMZ.

I could even whip out selinux policies that would let me know if the LLM tried to drop an executable where it's not allowed.

business genius: yes unhinged: also, yes

should he hold all the power he does? no, he's unhinged.

as you were already told BYD surpassed them, by quite a margin as well, and if you add the other chinese EV maker it's huge, even if you pit all the western EV car makers against the chinese one they are eating our lunch.

On top of that Tesla never reached the scale of any other car maker, those are some of the biggest and most complicated companies and supply chains on the planet.

I used to think we owed him the fact that EV are a thing, I then learned that China was preparing for it for a while, now I'm not so certain if he had the vision that it was possible to overcome the stagnation and resistance of traditional car makers or he just knew the chinese were also working on it.

for sure he saw the promise in reusable rockets, as he saw it with LLMs, a business genius he is, an industrialist I'm not sure.

what other sandbox do you need besides running a harness in a container with the right folder mounted and some due diligence like -nonewprivileges? are we talking host network sandboxes here? I think he's just referring to a rogue agent running rm -fdr --no-preserve-root, and that's safe in a container.

Talk to your Oracle Cloud representative wink. No discount on Nvidia, plenty on AMD. IBM is also offering incredible discounts on Intel Gaudi accelerators.

I even find that for what amounts to a web search Gemini 3.8 flash is better than the big models, stuff like the links to updated tax codes in different countries, or whatever MS is doing with Azure/o365 or AWS with their plethora of products.

I think use cases are the real reason why people have such different experiences, I too find that Opus5.5/Astra/6 are better for UI/UX now, it wasn't the case a year ago, at some point Gemini pro 3.5 was the best one at that.

That's also why I use all of them and try to not be locked to a single harness as well.

I have similar experience than you, with a big caveat, opus 5.5 would have still broken badly a custom ERP.

And on that specific kind of software, ultimately a big CRUD, there really isn't that much of a difference between GLM5.3 and Opus/OpenAI.

You see the differences when you get to different class of software.

I also have data entry applications that use LLM to actually parse documents, it's all Chinese models self hosted because the economic calculus beated a hosted API by about 5x

Because it's just Opus 5.5 and Haiku, and on OpenAI side Luna, that changed the calculus. Also, I find all of them, including GLM5.3 and DS4.1, to be 100% en par with American's frontiers model for CRUDS (90% of enterprise programming).

Tangentially, all of them would have broken quite badly custom ERPs from my own experience.

2000$ dollars per month will get you 8xAMD-MI300X on Oracle.

I have deployed multiple setups with 2/4/8 x H100/200 to do data entry with LLMs at big companies. Trillions of tokens already inferenced ok those. The starting price is about 100k.

She is. Those things are still hard. I run my own inference, I wore my own harness, I implemented production apps using LLM ("document intelligence", aka data entry), I followed Karpathy course and trained GPT2, I trained a classifier.

It's still hard for me to explain a lot of nuances to IT directors with a technical background.

They would leave a competitive advantage to non-US companies. Even if, and that's a big if, they can coordi ate with Europe Asia would still ignore them

Hosted in a cloud: a VPN Lowendtalk Is an amazing community to get offers for those.

At home: truenas and gitlab.

At home, for work: I have an older HP workstation with an older Xeon with 24 physical cores and it runs a rhel (I have them for free as a partner) and I use it to run VMs for experiments and replicating infras, K8S, Windows AD stuff, etc.

And yet my own value for the company has actually increased. And the LLM alone performed like an LLM from 1 year ago.

I'm not sure how this will translate to math, I suspect it will be the same.

I wrote this earlier. I am a sysadmin and do devops. I just spent 7 days writing what seems like a trivial endeavour with LLM, a wrapper around imapsync to migrate terabytes of email from an onprem server to exchange online. opus5.5 would have broken the exchange mailboxes at least four times.

there exist dedicated email migration software, it's pretty expensive, it was worth it for me to spend the time to learn things about how MAPI, IMAP and different email clients and different email servers handle things, their quirks and edge cases.

Before LLM it would have taken me months, not 7 days, to produce what I did, so the calculus would have skewed towards either buying the email migration software or other subpar solutions, accepting a lot more errors and losses during the migration to be handled with a different process.

With LLM it made sense for me to spend this time and, as I said, I learned quite a lot about emails.

I just had to move a bunch of emails from onprem to exchange online. there's software that already does this well, imapsync, I just needed a wrapper around it to automate it and get nice reports, then I realized there were quirks in how the origin server interacted with it's email clients (not outlook) and how this turns into edge cases and issues while syncing live mailboxes to exchange online (expunge, flags, runs where you get interrupted, cutover times, yadda yadda yadda).

it quickly turned into quite a mess, opus5.5, often on ultra, almost a week of work (that's 7 full days, not a calendar week), I still don't have very precise reports because of all the edge cases (5.5TB of emails across hundreds of mailboxes) and opus5.5 would have basically broken the mailboxes on exchange online at least four times. it's still worth to do it like this, because I keep strict controls over it, instead of buying very expensive email migration software.

same experience with using LLMs (which I do) to write critical code for what are basically CRUDs but actually handle invoicing, warehousing and all those nice critical problems where there's actually money on the line but are not truly deep CS issues.

again, it would have broken things quite a lot if I didn't understand the problems deeply.

then I ask it to implement what is a complex alghorithm for me, while explaining it to me, and I learn new things and improve a lot.

it's still a jagged frontier, I'm not sure how this will go with math

I am becoming senior, as of now I'm still getting pay raises and experience I'm cherishing, but you have to be strategic in this phase as well, arguably more?

We built a complete production-grade inference service from scratch on a cluster of more than 100,000 Chinese-made AI accelerators. All production inference for GLM-5.3-Flash runs on this system.

the fact that it's mostly about policies does not make the link list useless.

What kind of programming do you do? If it's enterprise programming and it's basically CRUDs all the way down, yeah, LLM are pretty good at it and you need to find a balance if you don't enjoy it anymore.

All this bitching does feel a little off to me though, I get it, but I don't think you guys get how bad it sounds from the outside.

LLMs are not perfect, even in enterprise programming. Even in my case, where a few CRUDs may manage 100s of millions of dollars, and the code LLMs produce is actually good enough, for sure comparable with mid developers, most of them, you do need to develop a system to keep everything on track, I've found my balance and honestly? It's good and I enjoy it.

Go into embedded critical systems, that's probably the last thing they will get good at.

Then they will leave the huge advantage in cost to the competition, I mean their customers competitors. Hard to fathom how US companies will not want to use the cheaper option when EU and Asian companies can.

I have deployed very successful LLM based software that reads sales people emails and inserts orders, or stuff akin to orders, in the rest of the systems. Can you write me a regex that parses a messy human email thread and produces a clean JSON with all the order details? It's been working for a year with less than 3% error rate, better than what the humans themselves were doing.

I remember an episode of TNG were Riker was instructing the computer to come up with an algorithm to search the surrounding space with certain sensor in a specific pattern, I don't remember the exact words but it was something like a sphere with growing radius with the enterprise at the center. we now, basically, have that, except for the space ship.