gmueckl

Is it, when the software has the potential to destroy one or more people's lives when things go wrong? Sometimes the devil you know is the far better choice.

The word "generally" contains a lot of assumptions in that first sentence. I can think of a couple of scenarios where fixing a bug can absolutely spiral. There are even cases where the far better outcome is to leave a bug in place now that it is known.

The trick is to understand what is going on here: rather than a singular model trained for this exact job, this is performed by a awful hackjob of two models cut open and glued together at the seam. That glue layer is a single matrix, a fully connected layer, that is the only thing re-trained in the combined model. The rest of the coefficients are fixed. Naively, this shouldn't work - the internal representations of each model are unconstrained and should intuitively have completely different shapes. But this paper shows that this isn't quite the case: the difference between the latent spaces can be mostly expressed in a single matrix transform. And that allows that merged model to work at all.

No, that cat is out of the bag. First, unilaterally forcing controls on models will not shut down open weights produced in other nations. Second, the capabilities are too powerful for military amd spy agencies to ignore. They will force the creation of unshackled models to attack IT systems, even if it must happen in secrecy.

Okay, why is it not good enough?

Mandatory tamperproof data retention for bookkeeping and taxes, GDPR, mandatory machine readable invoices; a flawed data backup routine can have considerable legal repercussions... it's the way those pieces interconnect that makes compliance time-consuming.

But they were not too far removed from how the standard Windows buttons looked. Office had its own look in a lot of ways back then, e.g. with the extra handles on the toolbars.

Now, if I take a couple of current applications, the differences are much bigger: from the top of my head, Explorer, Settings, Edge, Office, Steam, Spotify all have much more unique UIs.

Electron is atrocious, if you asked me. Modern computers are wasting staggering amounts of CPU cycles and memory on inefficient, bloated software stacks. But the realistic response to the RAM shortage will be to tell users to run their stuff in their normal browser instead because that gets rid of the extra browser runtime.

Is what I wrote confusing? I have nothing against labor laws in the EU. It's that there are a lot of other well meaning regulations that bog down young businesses. They are mostly justified, but they are also pretty daunting chores for founders.

No, labor laws are fine exactly as they are. Not importing US style at will hire and fire culture is an asset. That's all budgetable and relatively minor. If you are seeing a need to downsize your company early on, the fundamental mistakes already happened way earlier by choosing a path that was unsustainable. Such a company doesn't usually have a future.

The real question is the minimum temperature to make them respond. I hope that it's a much a lower setting than "Oh shit, the Russians are already marching towards Warsaw/Helsinki!"

This is getting priced out of the market for the foreseeable future: thick clients are getting squeezed out of the hardware market because rhe componemt prices are surging - the datacenter builders are trying to gobble it all up.

Meaningful local computer power could only survive realistically if exceptionally high demand (multiples of that of AI data centers) could be generated within weeks/months to get the semiconductor companies to pivot and sell the components mainly though other channels.

64GB laptops are no longer in stores. 32 GB is probably going to get squeezed out soon. For prebuilt desktops it's the same. That's no basis for meaningful local, decentralized compute.

No, this article gets its core premise wrong: labor laws are not the main issue. The issue is that if I wanted to found a new company, I'd be confronted with lots very strict regulations with lots of implications and induced overhead. The list is too long for a HN comment.

It's not impossible to overcome, but it is an added commitment for any new company. Founding is not very attractive and that's why we don't get that many crazy wild bets here (Tesla was one initially too).

You'd be surprised at how much is suddenly possible in the EU once politicians feel real fire under their butts.

When the end of a sprint is a hard cutoff point for tasks, agile is done wrong. Just carry unfinished work over - it won't go away, anyway.

Abandon the Paris agreement and then what? International politics won't allow for anything better to be put in its place for the foreseeable future.

That mostly changes what people worked on rather than the total amount of time spent working.

Well, the quote says that rhe first iteration was habd coded. So I assune that this means humans fully inchangé ofnthe décision-making; otherwise, the statement would be dishonest, wouldn't it?

I think your lawyers might have a word with you, too. Source code history is more than just a record for developers. It's also a historic record that can become relevant in legal disputes.

The humans decided they wanted that specific conversion tool. The tool just worked on creating it. The decision is the human creative input.

This also speaks to Bitcoin's design weakness in the form of the distributed, authority-less blockchain: there is no authority that can undo or correct illegally made transactions. I know that this is used as a selling point for crypto coins, but I would argue that it is the central enabler behind these conplex technical attacks because it shifts the economy far into the attacker's favor.

Germany deals with multiple thousands of WW2 bombs each year to this day.

Doesn't this ignore the costs for new masks?

Naive question: would flash or some kind of write once memory be dense enough to replace a mask ROM? I'm wondering whether it is feasible to manufacture "blank" chips at the semiconductor factory that have a fixed architecture, but are model-agnostic. These chips would get the then-current weights burned in on first use. It's probably quite wasteful, but it could keep the same chip design alive for a longer time.

The central room feels large today, but I bet it was the right size for 1930s technology.

The space requirement comes from two things: the diagrams on the panels are actually fairly complex with lots of controls thwt fit mostly inside the diagrams themselves. Operators in the room should be able to scan the whole diagram easily from a central position.

Thw size of the room then becomes a requirement because all the panels need to be oriented inwards, not occluded and they can't shrink down too much because the controls need to fit.

Why not? Friction from tidal forces keeps the moon got enough to contain liquid water under its ice cover. This makes it similar to Europa in some ways.

Models can't learn from misbehavior after training. Any session is an independent context and there is no mode for punishment or deterrence in production.

Corrective punishment in the real world relies on the receiver's rational and emotional responses as well as their ability to remember that episode. Even animals respond to such treatment. None of these levers exist for ussrs of LLMs.

A terminally cynical mind might insinuate here that focusing on the product is a way for AI companies to keep doing their own business as usual, no matter how negligent that may be.

Is it? A lot of software is really just a pile of context-specific database front-ends. There is virtually nothing difficult in the implementation itself. Most businesses in this space survive because they know how to dress up their pile of CRUD for their customers. And it gets worse: a less reliable, buggier product may still be the better fit for a customer if it works well enough for what they use it the most.

I also hate preachy sermons about new tech with a deep passion and that transfers somewhat to the tech they preach about.

Despite this, there is a reality that LLMs for coding can create pretty decent results fast. There is something real and usable hidden in all that mess. It's like the dot com bubble in some ways: the bubble died with all the overhyped "x on the web" BS that had no real substance, but the web itself stayed. This time around, LLMs and agents will stick around as tools, but the hype will likely die a loud, messy death.