No reason to expect Moore's observation to apply there (though, maybe?), but it will have big implications for power usage.
1,560 karma · joined September 15, 2023
No reason to expect Moore's observation to apply there (though, maybe?), but it will have big implications for power usage.
All of that will all be legacy in a couple of years. Today's B200 clusters are tomorrow's e-waste. Decentralization might happen gradually or abruptly. But to me it's obvious that we'll be thinking of high-tech tensor processors and GPUs the way we thought of individual transistors and tube amplifiers in the 1980s.
If AI turns out to be the revolution it purports to be, than the underlying hardware will change much more rapidly than it did with ICs and microprocessors in the late 1970s. Today's hot is tomorrow's junk.
- do not get on a cruise ship
- do not get off at a remote island
Access to these visuals really helps understanding what they are talking about
Escorting them to the door, and revoking access for the remainder of contract yet paying wages for that period seems very descent. Off course, you don't do that when the termination was triggered by employee's misbehaviour.
But, yeah - the point I was trying to make is that there is only so much you can do as an employer to protect the company while there's an infinite number of reasons for anyone to be traumatized or otherwise act erratic. Admins are always entrusted with huge power and while wariness is probably warranted, distrustfulness is IMO counterproductive and often harmful.
Added, later: hey you changed your comment, added a whole paragraph.
Chrome..., Pixel..., Gemini..., Android..., anything but Google, please.
But "sharing data" does somewhat suggest they are sharing the payload, the actual conversations.
Then came the model trainers, ignoring the entire discourse, reasoning: "if I can download it, it's mine too use". And then basically selling the resulting tech back to the community.
Not unlike big tech extracting money from open source, but at least the latter usually (somewhat maliciously) complied with the license.
But it turn out surveillance works just fine if you only focus on the meta data. Knowing who takes to whom, and which sites people visit is much more valuable (and much cheaper) than scanning the actual payload.
And why collect all that data yourself if ad companies are happy to sell it to you, ie to the government? (Huh, maybe that's why Facebook changed its name to Meta, come to think of it)
First and foremost to me those screenshots are somewhat disappointing as they can't match my memories. NeXT, BeOS, Irix, OpenLook, SunOS, Arthur (imagine the diversity)... they were SO awesomely impressive at insanely high multi-sync CRT resolution.
Reality simply can't match the mind's eye, at least not for me.
Let's get rid of those pesky ACLs. Trust us, we know what we're doing.
I'm hesitant to admit even that. Like any computational linguistics problem, accuracy relies on coverages of all levels: form morphology, through syntax and semantics to speech act and world knowledge.
I worked with state of art speech recognition in healthcare setting. The model was specifically trained on small set of languages with emphasis on covering medical terminology.
It worked great for conversations most of the time, but sometimes messed up very badly. For instance when patient would mention the name of a relative, a street address or phone number. Spelling out an email address would mess it up completely.
It's just like when you're a horrible typist and rely on spell checking: The red squibles are gone, but the story no longer makes sense. Or when you "autofix" a syntax error, but the meaning diverges from your intention.
As the technology improved the number of words decreases, but the mistakes get more severe.
It took years of activism and voting with our money to get banks, pension funds and similar institutions to stop funding cluster munitions, land mines, nukes, oil, tabacco. Now big tech and some AI companies are on the radar.
I predict the B200 data centers we're build today will be obsolete in 3 years and we'll be using whatever models and hardware that isn't even on a road map today. Likely not NVIDIA, likely not OpenAI or Anthropic. Maybe Chinese?
In the mean time, we must continue building software with the clumsy coding agents tied to cloud services as this (for now) seems to be about the only area where AI economically makes sense.
...But there is always at least one hacker.
The issue with hardening DRM is that at the core it's hard to protect against an adversary that with physical access to the device that keeps the very secret. From the vendor perspective, the very customer paying you is your potential enemy.
That means that the root of trust isn't itself protected with cryptography. Instead, it relies on security-through-obscurity, Faraday cages, fuses, anti-tampering and lots of glue. And it's a numbers game if there are thousands of different devices, potentially with different flaws while your adversaries are hidden among billions of customers.
There is still a gap between the hacker and main-stream availability, though: laws and legalism, like DMCA that penalize disclosing how the obfuscation and all work.
It won't matter to the masses, it won't hamper "bad actors" because hackers will find flaws instantly.
It's just enshitfication.
I suppose that most of all, it reminds me of time when actual, genuine real innovation in UI design was still on the menu.
For instance, sprinkling a bunch of nodes + sensors in hostile territory should allow for gathering intelligence, guiding drones, setting of fuses...
Imagine these 8 bitters were mostly hardwired, with less than a million transistors.
* with internet access to FOSS via sourceforge and github we got an abundance of building blocks
* with central repositories like CPAN, npm, pip, cargo and docker those building blocks became trivially easy to use
Then LLMs and agents added velocity to building apps and producing yet more components, feeding back into the dependency chain. Worse: new code with unattributed reuse of questionable patterns found in unknowable versions of existing libraries. That is, implicit dependencies on fragments multitude of packages.
This may all end well ultimately, but we're definitely in for a bumpy ride.
I'm thinking of the first episode of Alex Garland minisery Devs, where a somewhat brilliant Russian guy demos predicting the exact movements of a bacteria. That would be proof that the model works. And even it it were to mispredict, most of the time, it sounds plausible.
And then you run into similar issues as the llm does, like silent failures, loops, contradictions unless you're very careful.
The essence might be the same closed world assumption problem. In llm case this manifests as hallucination rather that admitting it does not know.
Nitpick: off course we don't really lack a theory of consciousness. It's just that Alan Turing choose to ignore all the existing prior discourse in humanities and philosophy.