I only skimmed it but the conclusion was it’s better value to rent compute by the hour unless you have several people using your local compute and/or have you such sensitive data that it can’t go off prem.
How accurate the article was I don’t know but it was an interesting point to consider.
I love the idea of getting a super powerful Mac mini and being able to run local models independently - mainly, you know, for the coming apocalypse - but at the same time I just don’t need to spend that money because apart from the occasional overnight run my AI is only in harness for maybe five hours each day and the rest of the time is off snoozing and grazing and eating daisies or whatever they do when you remove the harness. So if I did want to go down the open model route use rent compute. Which I get doesn’t quite solve the problem you’re trying to solve which is end to end ownership and control.
The danger of AI models isn't a rouge agent or user, it's that it ingests all knowledge, art, and other human endeavor and then restricts access to it.
Closed models are unethical.
Our benevolent overlords just decided they would give us a taste before running off with the goods. Also, our jobs and all those rare books they destroyed and goodness knows what else...
What matters more is that LLMs were popularized and scrutinized to the level that was necessary for people to move on already. We're far from done with AI, but LLMs brought a previously indifferent public up to speed with politics, philosophy, economics, etc.
We don't have more compute with cloud AI. All we have are perpetually rented gpus. The only way to get anyone to pay to rent gpus was to make pc hardware too expensive to buy. That is what jensen did. Jensen colluded with other tech CEOs to raise prices to create artificial scarcity by buying gpus and ram they did not need.
How else are you going to get the terabyte+ of interconnected VRAM needed to run anything remotely frontier?
There is no reason why these high speed interconnects would not filter down to consumer products. The only reason prices are not falling is because of the market collusion to keep prices high.
Right now its $2500 dollars for 4tb nvme raid and a pci-e switch capable of connecting a graphics card to that memory at 60GB/s. This is stuff you can buy during an artificial shortage.
I would have expected cheap 8tb nvme raid cards capable of 120GB/s to be available with the needed pci-e switching for less than that. Collusion denied it to you.
I think that, through cold and cynical lens of commercial self-aggrandizement, the AI Safeguards are merely the _good_ reason (e.g. we are protecting the consumer (and the world!) from our awe-worthy AI, and from evildoers who seek to use it for their evil ends), while the _real_ reason is much more predictable, boring, and disappointing: AI Up-Selling (e.g. our freely/cheaply-priced models (which nevertheless are very, very _expensive_ cost-wise) are already corrosive and damaging, and are regularly used by nefarious evildoers all over the world (including ourselves!) but we want to (no, we are compelled to) find ways to extract more money from _all_ of you, even if it means lobotomizing the models, so that they burn more tokens to solve a problem (in the best case) or to not solve a problem (worst case for you, best case for us, because now you have to ask us for the more capable, super secret, super exclusive models, that will, invariably, cost whatever you can afford to pay, to make the pain and insanity stop).
Or you silently get downgraded to a dumb model, while still being charged the higher model, and loads of "thought" tokens burnt to go "lol pay more sucker".
Shouldn't be, but is.
Libraries might become popular again if not seen as suspect of supporting subversion of the system.
So does Wikipedia, more lightly
Felony contempt of a tech company's business model.
Stabbed his brother Kurn in the chest
Sabotaged Risa’s weather-control system.
Helped Riker fight local guards to free a detainee.
Authorized the erasure of Kurn’s memory.
Killed Duras to avenge K’Ehleyr.
Brawled with fellow officers during the strike at Quark’s.
LtWorf, it's official, your namesake has better moral taste than you have.So I did a little more digging and found the flag I needed to put in a config file to bypass forced updates. I went back to the LLM and explained the workaround and it acknowledged that is was aware of that information the whole time. I forget what it said when I asked why it didn't share that info originally, but it was probably some ole BS...
What is our solution? Relying on Chinese open-weight models will be unreliable, to say the least. No one will invest another frontier lab like Anthropic or OAI, let alone one that will release the weights. Unless the technology improves like how workstations outcompete mainframes so the open-source community can build our own models, I fail to see how we change the situation.
NVIDIA, on the other hand, released a huge training corpus and open models up to the 500B class.
What do you mean? I've only used them for writing code and they seem to work well enough for that.
Now ask yourself: What else have I not inspected in this model?
The Chinese models have no technical advantage over open-source, open-weights models of known provenance.
Particularly when the thing in question is completely lawful, benign, and when it would be unlawful for the state to restrict you from doing so.
And AI companies like Anthropic want to restrict your ability to use alternatives through regulatory capture effectively boosting their capacious and unaccountable censorship to sate level censorship while bypassing all the legal and social restrictions on state level censorship which you'd otherwise enjoy.
At this point it's pulling the lever and the machine maybe says no and refuses to do to the work, and it is getting worse. I'm basically paying for a toy because no one can actually depend on that.
We need affordable and freedom respecting AI capable computer hardware in every single home. Allowing them to concentrate compute in their data centers is an existential danger.
The closed/paid models can only get more restricted as time goes on and they're going to be a dead end if you want to do anything outside the norm.
I completely refuse to accept a computer telling me I'm not allowed to do something, so I've been increasingly using open, self-hosted models.