Well the 'nice' thing about this sort of language is it can frequently be true multiple times. It's "better" but how close to "good enough" is unspecified.

The Anthropic one was interesting because the original person behind it was fairly nuanced and honest. The stunt needed an existing reference implementation as a basis as well as a boat load of unit tests and needed hand holding and still didn't quite pass the big test of compiling the kernel (needed to borrow missing bits that claude couldn't figure out). Then as many noted the compiler kind of sucked, compiled code that should have errorred, failed to compile code that should have. It was interesting that it could keep iterating, but the end game in the heavily tilted game remained elusive. The original post was pretty honest about this. Then an executive says "this says software development is already dead!" when the experiment was pretty much exactly the opposite, that even with the most capable LLM available, it needed a *lot* of human development aid to get to end game, even as it was given something that already worked to knock off *and* had certainly trained on multiple C compiler codebases.

Yes, but before, they wouldn't get greenlit and projects with a vague chance of progression would.

Now they precluded quite a few projects that actually could have progressed.

I tried to extract specifics to expose the ill conceived ones, but the other executives on the panel were still a fan of the slop ones, and preferred them over more grounded.

To make that, first we need a court to actually rule that 'training' can count that way. Currently everyone is operating under the philosophy that ingest to an LLM "doesn't count" somehow.

Well, no, that assumes exactly one implementation for a given feature in the wild.

Imagine generating a random string. Hundreds of codebases will have that same function. So this process may pull that from any of those codebases and not necessarily from the source codebase.

It's never generating fundamentally novel code, but it is drawing from a huge training data that includes the same thing done dozens or hundreds of times with technically distinct code.

IBM BIOS wasn't source available either. The precedent for 'clean room' involved reverse engineering binary code.

So while the current story emphasizes the loss of open source protections, the same principles would apply to LLM transforming binary and test cases to a specification.

But it's meant as a proof point of our current interpretation of LLM and copyright. So far this "counts" as clean room because courts have not said LLM ingest is a violation, and they are using the LLM to launder the code to an intermediate form and then to code based on the 'clean room' finding.

So while you are right in a sense, the point is from a court perspective this is "equivalent" to clean room unless new laws/court cases amend the status quo.

Anecdote to back this up, we have annual round of employee directed projects where people propose something they will do that no one asked for in hopes that they do something unexpected that's worthwhile. Generally it's a waste of time business wise, but at least people get to work on something they actually believe in.

Anyway, usually they at least usually manage to create a somewhat working demo of their concept, but this year most of them failed to do so, because most of the pitches were people that didn't know how to do the work, but GenAI was able to generate pitch material that convinced executives to approve them and largely drowned out the people with actionable proposals. So most of the final presentations were people just repeating their pitch and hoping people didn't notice they had no new material since their pitch a few months back.

Quite frankly a huge amount of skepticism is absolutely warranted. The AI tech companies are broadly worrisome, but Palantir takes the cake for outright villainous efforts.

To the extent they have shown ambition for a future, they haven't shown they have a whiff of folk's best interests at heart.

Nope...

Note the intent to "retrain" labor.

The AI dividend in their scenarios wawould be a trivial gimmick. They still want the labor force toiling, but a dividend to mollify concerns about AI displacement.

If this happened, I would bet maybe 100 or 200 dollars a month of "dividend". You'd still be expect to toil away under the capitalist rules to actually have a credible living.

In terms of what to change to doing, some of these folks already said that people need to return to manual labor. They really hope AI will work as a strategy to make educational an impractical choice and people just kind of stay uneducated and desperate to provide manual labor for sustenance.

This is a path for them to patch what they see as a problem in capitalism: some modicum of class mobility. Other than that possibility, capitalism works great for them. For communism as a core principle, then they need to go full authoritarian, and the power struggle among them to get there is a more dangerous one than a modified capitalism.

Ok, but how does it work in the scenario where it pushed another package into the street and then the package gets hit by a car? A package that may have nothing to do with Amazon aside from it being pushed into the street.

Also, from my experience with this, it's true that 90% of the time they are pretty agreeable, but coincidentally the 10% of the time where they are skeptical just happens to be the more expensive stuff.

I just don't order expensive stuff from amazon. Actually, Amazon is only if I can't get it anywhere else reasonable nowadays.

Of course, problem is the lack of availability of similar data for successful companies.

To the extent it *might* work, that's all the stronger case to keep your successful company's information confidential, to avoid helping future competition be a broadly more capable company.

So we are still stuck at training your model to be a crappy business.

Train models mostly on the communications of failed businesses... What could go wrong?

Think this was more about the business side than the execution side.

They can be just fine if Linux desktop is seen as 'acceptable' in the mass market. The operative word being 'seen', not if it should be acceptable, but if people believe it to be.

It's a tall order to shift the perception of the mass market.

If Microsoft screws over their users, and now 'just enough laptop' can be bought from Apple within a price range long deemed 'adequate', then AMD/Intel essentially *need* people to decide they *love* linux desktops, and broadly speaking at this point the mass market is "a device that runs a browser and who cares about OS as long as it's not actively pissing me off".

It actually might go badly for Linux desktops if Windows screws up their market share too badly. Apple does not give a shit about Linux support and so Linux desktop largely lives on the standardized ecosystem in x86 side in part thanks to the separation of concerns between the hardware vendors and Microsoft. If Microsoft managed to kill the Windows desktop hypothetically, might be hard pressed to have any hardware for Linux desktops to run on anymore...

Because GenAI screws up, and screwing up is less of a big deal for the attacker, but can be a huge deal for the defender.

Attacking GenAI fails and either your attempt does nothing and you are no worse off, or it accidentally trashes a system that you were trying to control or copy data from, but you didn't care about that

Well, compared to the Windows release strategy...