foobiekr

Working for a vendor that ships Linux code, we find regressions all the time. Because we actually test things. We do try to be good citizens, and Linux is the right answer, but it isn't a false assertion on the part of the author at all.

In a previous job, 20 years ago, we found that the Linux USB stack was so bad that it was full of deadlocks and unprotected shared memory on 2.6. Just atrocious. And subsequent releases broke it again and again.

About 5 years after that, working for a different company, I had a well-known kernel developer tell me point blank that as long as we release our code to the kernel, they will "keep it working." When I probed what that meant, he genuinely saw no difference between "compiling" and "working."

There are various efforts to add testing/tests. Those are good. But we lost a lot when Solaris, Irix, etc. all died, and Linux has never quite gotten there, and I doubt it ever will.

OpenAI isn't making money. They are like a giant furnace for investment dollars. Even putting aside that the NFL and OpenAI aren't the same kind of entity, there is also no taxes issue.

I don't understand what theoretical basis can even exist for "I don't know" from an LLM, just based on how they work.

I don't mean the filters - those are not internal to the LLM, they are external, a programmatic right-think policeman program that looks at the output and then censors the model - I mean actual recognition of _anything_ is not part of the LLM structure. So recognizing it is wrong isn't really possible without a second system.

If you look at any map of antibiotic resistance, it is quite obvious where the problem is.

You can tell. The tiresome lists.

It is not an issue in the US.

It is a massive, massive issue in some very large countries.

Does ollama still execute whatever arbitrary python code is in the model?

Given the delivery mechanism for OpenAI, how do they actually keep it private?

That’s going beyond simple use. The two year lookback is real but nobody is going to fail a basic clearance over college pot use.

Silly and convenient. No one takes the claim that you can’t measure software productivity seriously and everyone simultaneously agrees that simple scalar metrics often fail to show the big picture in any and all disciplines.

The rest is just usual software guy hubris and lack of awareness of the discipline.

Revenue per employee $ spent.

past drug use does not preclude a clearance (edited)

So much confidence about how this thing we’ve never encountered will not be a problem.

I was not aware that my last two companies have secretly been Intel because every one of these factors applied to them.

It’s really another aspect of the innovator’s dilemma - interesting $-wise and margin-wise is so limited that they end up never doing anything interesting.

It is a problem really specific to high margin companies.

You actually don’t see this behavior at low margin businesses because the spectrum of things that can be done which aren’t immediately labeled a margin drag is a lot wider.

The company I work for now put all the up front effort into a new product line which would lilekk pk y have easily gotten to a billion or two a year within 3y with the right customer wins, but out normal margin is very high and the new line of business would have been very profitable objectively but would have been a margin drag. Every dollar would have lowered the gross and net margin.

Hans Moravec had a suggestion on how to do this in Mind Children: instead of examining the whole brain, you measure a layer, then replace it, layer by layer, until the last. There is never a copy and an original, it's just a ship of theseus self where the neurons are individually replaced with new ones, albeit electronic ones.

Correct for inference - the main use of the interconnect is RDMA requests between GPUs to fit models that wouldn't otherwise fit.

Not really correct for training - training has a lot of all-to-all problems, so hierarchical reduction is useful but doesn't really solve the incast problem - Nvlink _bandwidth_ is less of an issue than perhaps the SHARP functions in the NVLink switch ASICs.

So many prompts blocked. ".. inspired by .." three out of four are blocked.

Depending on how he managed things, possibly tax advantaged; especially if he could take advantage of the Qualified Small Business Stock (QSBS) tax exemption - 100% capital gains with no taxes.

It's not even in the same ballpark. With excel, you are inspecting _the output_ for a specific run. There is no equivalent in modern programming.

Spreadsheets let you click any field and understand where it came from. Excel can be thought of as an amazing post-execution debugger where you can get a previous output and value for almost any piece of memory ("cell"). I would love a debugger that could do this post-execution for some output.

Only if the thing consuming it can consume formulas.

These days, I prefer google sheets for most things spreadsheet, but Excel is far and away more capable, and after a certain level of complexity, just unbelievably fast in comparison. Complex sheets in Google Sheets get bogged down very quickly, and the data-access support is weaker and handles errors very poorly.

neither of those are useful for the next stage of analysis.

Stop doing this. You are confusing the perfectly noble aspect of calling it abuse material to make it victim centric with denying the basic purpose of the material. The people who worked hard to get it called CSAM do not deny that it’s pornography for its users.

The distinction you went on to make was necessary specifically for this reason.

It doesn’t. You’re kind of illustrating what I’m getting at.

It’s not “ridiculously low.” Before the layoff they were 2700 people. That’s like a VP2 at most large tech companies and 1.5M total comp is spot on for that level.

The most likely explanation is that the code you are writing has low information density and is stringing things together the same way many existing apps have already done.

That isn’t a judgement but trying to use the ai code completion tools for complex systems tasks is almost always a disaster.

I am a huge Amiga fan, but the Amiga was going nowhere and was never going to win. The OS is just as terrible as classic Windows and MacOS from a reliability standpoint; yes not using a message pump for timeslicing was a really nice property but in most ways the design was _worse_ in terms of any hope of eventually getting memory protection in place.

I love the Amiga - it represented a unique point in time that coalesced a lot of interesting technologies and people together trying to do something interesting - but it was as far from a technology that had long term potential as you could get, pretty much in every way.

I wonder how much impact there will be on average life expectancy uncovering that a lot of the super-long-lived population examples have been pension fraud.

You can never roll your own and have a working ecosystem. This is exactly the problem.