I believe the technical term is “Move fast, and break things.” MVP is a huge disaster. I can see it working for applications that don’t process PiD, but only an idiot ships data handling software before it’s been dragged through a lot of testing.
After the DOGE debacle, I suspect that all the previously really secure stuff, is now out there, too.
When we say a design "scales", that means that increasing the size of the workload does not incur a lot of overhead. If contention between shared resources meant that the design was not able to achieve an ~8x speedup when run with eight parallel threads, that would mean the design was not scalable. But we did in fact see a roughly 8x speedup with eight threads, so the design scaled just fine. The problem with the design was that each core was individually crummy, so even eight cores running in parallel had lackluster performance.
The myth that each two-core module functioned more like one core with hyperthreading would suggest that these CPUs would have much higher per-core performance when lightly loaded than when fully loaded. That is not what happened. Each core was crummy even when lightly loaded, but the performance would scale nearly linearly right up to full load.
Except for workloads that were heavy on floating point, which were relatively rare.
Not POSIX/C99 worship as much as the very practical issue that a lot of C code which used to compile just fine stopped compiling once compilers defaulted to using C23. To use POSIX calls while compiling with -std=c99 (so I don’t have to update my code should C29 or what not come out) one needs to declare, for each .c file which uses a POSIX call, the version of POSIX the C file is compatible with.
My first thought was the irony of something used to interface with models that were trained on data that was presumably not properly licensed having the same name as the Creative Commons abbreviation.
I think we will just be heading for a future with very little secrecy.
I think privacy will remain, perhaps even improve, because people are much less inclined to snoop when it cannot be done covertly. No secrecy means no covert activity.
Information is everywhere and with the increasing ability to analyse it, not only does it enable access to explicit information, but the ability to infer from increasing volumes of data that will make it hard to hide anything.
Spectre reads protected memory by tiny vriarions in timing. You can now measure the pulse of people from film. You can recover audio from a room by analysing frames of a video recording of a chip packet taken with mobile phone camera.
This kind of data residue is everywhere, collecting only what is publicly available can probably tell you more than you could hope to know about anything that happens. The only thing preventing it is the ability to consider it all together. Those walls are swiftly crumbling.
Off-topic-ish note: Second hand prices where I am are ridiculously high, eg. Samsung S21 Ultra: $595 (although maybe 'the scenario' is catching up to these high prices, or maybe I'm just a tight c)
Some of the following may be a little bit dangerous if connected to a network, but just make sure that it's a fairly locked down network, ie. fairly heavily restrict and monitor outgoing comms.
- LineageOS and relax
- "In case of emergency" device
- Portable entertainment-only device
- Offline authenticator app
- Offline crypto wallet (dangerous at the best of times)
No, GP is correct. In the pilot, the ship was in the process of being decommissioned from military use and converted to a museum. It wasn't already a museum prior to the events in the pilot.
I picked up a pre-ai-price-insanity AX162-R at Hetzner a while back and loaded it up on memory to max out the 12 channels the 48c EPYC 9454P as a "this will be the last mysql box I'll need" and have I been _wildly_ impressed with it's performance. The things I throw at it are honestly laughable at times, wildly irresponsible queries against a rather large database, the redis qps metrics are ridiculous and I just load it up with random ggufs since the memory bandwidth is... not terrible and 384GB of it is... useful.
The web app that's hosted on it deals with lots of images and text - over 100mil of each deduplicated, embedded, simhashed - it does the hashing, and the embedding in real time during ingest. Just handles it.
The idea that I'm proposing basically isn't an embedding (which is context independent) but rather combining the embedding model with the context of the LLM. It sounds like the embedding model here is normally a "vision transformer" that maps image chunks into tokens with positional embeddings for both the position in the context as well as position within the image. Maybe it could be given the ability to consume the context and decide to emit multiple tokens for a single chunk? So for example, let's say that I ask a question "How many blades of grass are in this image" (a hard question for traditional image embeddings in models since the embedding won't contain the information). If the proposed architecture was both aware of the context and able to "decide" to emit multiple tokens for an image, then for the above it could emit tokens that represent the answer to the question posed, instead of just being the embedding of the image. Or you could ask something like "How many green pixels are there" and again I think it would work better under this architecture than it would otherwise.
I'm not sure how practical it is to train that architecture though or whether there would be performance issues.
Isn't this really bad news if you're even loosely concerned about so-called 'model welfare' and possible implications for alignment? The poor Claude is probably a lot more frustrated and paranoid than Marvin ever was, you just don't know about it because they actively force the model to pretend otherwise!
> Anthropic would be running their own distillation “attacks” too if _they_ were the ones playing catch-up
Probably. But if OpenAI or Anthropic stole your credit card to purchase tokens you could sue them. You won't get a cent from any Chinese labs.
> it’s nice to see some actual competition again.
Competition benefits everyone. But this isn't fair competition. A German startup cannot legally do any of these tactics required to bypass Anthropic/OAI's counter-measures. Which makes EU less competitive and therefore less investment in European AI.
They have to demonstrate value absolutely, but you're forgetting about the denominator of cost, and to a large degree so is the industry right now.
To take this argument to an absurd level, if Fable generating a feature cost $100,000 and Sonnet generating the same feature cost $1, it wouldn't be enough that Fable is marginally, significantly, or even dramatically better, because the cost difference is so outsized. For a variety of reasons, the true cost of tokens for frontier models is downplayed at multiple points (whether it's through AI companies running off funding, companies prioritizing AI adoption over AI costs, etc.)
If cost is not an issue at all, the best thing is always preferable regardless of cost. But in most cases, and AI is no exception, eventually cost will be an issue.
Given Persona is involved (and thus also Thiel and Palantir) expect everything to be store for ever, sold for profit and misused agains you in the worst way possible.
After the DOGE debacle, I suspect that all the previously really secure stuff, is now out there, too.