The default state of agents and LLMs is inert. It requires action from a human even be able to do something. From the very basics like starting the software, connecting it to a network, having hardware to run on.
But the most important wrt AI is to keep the owners/operators responsible. Don't let them weasel their way out of it. They are for sure trying, and will continue to. This includes using language to overemphasise agents importance in bad outcomes, in order to downplay their own responsibility.
Agreed. AI as an "accountability sink" is an incredibly bad idea. It allows/incentives bad actors to do bad shit and get away with it. Which will, generally, tend in such practice becoming more common. Everyone loses except for the crooks.
We cannot accept "AI" absolving humans of responsibility.
Some kind of compute-in-memory architecture is a good candidate, I think. There are many alternatives here, researched for many years prior to the LLM craze. However economies of scale dominate in chip industries, and this tends to favor more conventional or incremental approaches (to piggyback on existing scale). Alternatively someone needs to have a way of bootstrapping the insane scales needed to be competitive with a better-but-different approach.
So it could be that boring and straightforward stuff like two-chip prefill+decode takes most.
I am missing a mention of ROM in the article. Keeping read-only weights in RAM is rather wasteful, as ROM can be implemented more cheaply. Approaches like High Bandwidth Flash (HBF) are relevant here, and should come to market in a few years.
Further optimization may be possible by tailoring for sequential access, since inference of a particular model is very predictable.
If you have good feedback signals, like tests/benchmarks/etc, then it is potentially better to do multiple turns where model uses that to adjust code. Which might not need as smart a model.
If Simon would pitch for example PCBWay that and I am pretty sure they will sponsor it (assuming their logo stays). They can do laser engraved versions also ;)
It depends on your tolerance level for having less than frontier LLM capabilities. First level worth trying, Qwen 3.6 35B A3B with a 16 GB VRAM (example 1x 5060ti 16gb, 600 USD for the card) with partial GPU offloading. Next level would be Qwen 3.6 27B / new Muse Spark / Gemma 31B with 32 GB VRAM (2x 5060ti or 1x 9700 Pro). Third level would be DeepSeek V4 Flash with 192 GB VRAM (2x Strix Halo at some 8000 USD total). These models can be tried on OpenRuouter etc, or you can deploy vLLM on rented GPUs to get a feel for what level you would want before committing to buying hardware.
AI seems to be on the same trajectory? Search was very useful in the start also, until it became entrenched. Then search placement became a target, and they are just focusing on extracting rents. All way paying the content providers zero or near-zero. And with years of that dynamic, we end up where we are now. It was the same with "social media". The same will happen with AI. AI is a power for more enshittification - being currently less shit than Google is (mosy likely) temporary.
And it is not success compared to another possibility, earning enough money to not need further fundraising.
Of course some businesses are more capital intensive and have longer time-to-money timelines due to factors outside the control of the company, and in that case fundraising is an essential tool for a long time.
Suck at what aspect of (analog) electronics specifically? Not contradicting the claim, just want to understand it.
I have not tested yet, but I suspect that LLMs with a harness that can execute code can do SPICE simulations rather ok these days? I have seen MCPs for measurement equipment also, maybe they can even close the physical loop?
The cards are also coming... Some of the very first are already out - but it will take a couple of years before they hit their stride in full. And cards will likely be primarily for Chinese companies initially, and then go for exports later. So it will go under the radar for many western techies for a while. Seeing what China has managed with PCB production, low to mid-end semiconductors, phone manufacturing and electric I would hesitate to bet against them.
Nvidia is now lobbying to remove the export ban, to try to reduce the incentives for developing their own high-end LLM training/serving (and cash in on current demand).
Even in a fine tuning scenario, the preferred form for making modifications is the weights plus the training pipelines, including evaluation protocols, tooling, etc.
Not comparable today. But in 10 or 20 years the situation might be considerably different. Hopefully closer to a everyone-can-build-using-hardware-they-already-have.
Does anyone have a good article on the topic? This one seems rather LLM generated. Not sure the specifics can be trusted... Overall thesis is maybe right, but would like to understand this in a bit more detail.
Capability per GB and per watt has also been going up lot. This will continue in the future as well (not necessary as the same rate as last years). But enough that I think Opus 4.8 level is reachable on consumer PCs within 10 years from its release. Say at the price point of 2000 USD in 2025 dollars.
Yeah China has a huge (and growing) advantage in power generation. And they have been looking to break into high-end chip manufacturing. The latter has high cost of entry and needs large scale to become viable. AI inference on own hardware would allow them to bootstrap the chip demand. Both for memory and accelerators.
Writing a Python extension would a good way to dip your toes into Rust, and also add a useful skill to Python programming (writing performant extensions). PyO3 is the main project, and the topic has been covered in several talks at Python conferences (check Youtube).
There are no limits to what you can imagine yourself hearing... And appreciation is a extremely complex subjective feeling, one could argue most of that is way beyond audio and acoustics, and even beyond psychacoustics into plain psychology.
When talking about the sensation, it is hard to differentiate between "really hear" versus "tricking oneself" - for most intents and purposes this is one and the same..
On the other hand, one can conduct blind tests that show that many phenomena cannot be reliably differentiated by a listener. Different listeners also have different levels of ability to discriminate sounds and sound quality. So when testing audio equipment using listening tests, one needs to consider the panel of listeners that one uses. Typically there are trained panels of listeners (who have received basic training and shown statistically an ok ability to discriminate) and "consumer" panels which are just random people off the street. The two groups will give very different ratings -especially for "medium" sound quality equipment.
For info on the former, see any decent book on psychoacoustics. And for the latter see for example Sensory Evaluation of Sound (Nick Zacharov et al).
I believe the proposed system is to run the containers on dedicated rented server(s). Instead of having the containers/VMs share underlying CPU/RAM with others.
GB200 and GB300 are decent for training LLMs? H100 can be used also? Though the current European deployments are rather small, even the biggest are just some thousand GPUs.
You also need a shitton of data. People are saying that synthetic data is widely used, but pretty sure that is on top of the organic data. Both for pretraining and posttraining.
That said, I do think we will see more open and collaborative approaches over time.