They are new proofs and for sure useful but as far as I’ve understood mainly interpolative. E.g. an LLM can “create” a poem about a purple chicken as it has datapoints for “purple” and “chicken”, so it can create something plausible inbetween.
Similarly, in my mind it can interpolate proofs by interpolating between data points for technique A and technique B. This is novel and brute-forcing proofs this way is useful. It is analogus to how sometimes it can generate programs that pass unit tests, I think.
However, creating fundamentally new concepts outside of the interpolated datapoints is not something I am convinced of. Maybe it can extrapolate some things, if correct add it as a data point, continue. Essentially a search, and it would be amazing if this works and maybe we can get some recursive improvement this way. But the “ideas” it will use to conduct this search are a function of the input data points as well, and thus in my view fundamentally limited in novelty.
Of course nobody can know yet really and I am just speculating just like you. But I also think the “experts” Sam and Dario also don’t m ow.
Ah, yes, Anthropic, the company who just got regulated and has an ongoing open dispute with the government, wants to regulate other companies with their copious political capital.
I'm still super excited for this release. KDE 6.4.3 from 6.2.5 is a nice jump. I'm very excited to get Wayland here, which is my normal baseline.
Maybe this time I'll try to get Niri running on this desktop. This is one of my daily driver systems, with a monitor plugged in. I'm typing on it now. I've been holed up on on 3.7.14, build id 20250701.1, since I don't want to lose the desktop, but this one seems worth losing my desktop for. Nothing but respect for Valve for working on ruac, a very nice A/B system image switcher, that powers all these updates, even if it means I'm about to lose this desktop. https://github.com/rauc/rauc
That's precisely where medicine is headed: personalized medicine.
You [hopefully] won't have to become a rare missed diagnosis because you didn't fit the demographic for this or that screening test.
Cost of genomic analysis is exponentially decreasing, and so much progress is happening so quickly.
Consider for example how in cardiology we advanced from ASCVD's 10-yr prognosis, to the PREVENT 30-yr prognosis. And still most providers are using the ASCVD score for their patients.
This was my experience too. I asked fable to implement a (quite complex, novel) CRDT engine. It did fantastic work for the 3 days I had access to it. The spec it wrote is exactly what I want. It used prototypes and examples to figure out some hard problems and answer a lot of complex design questions. Claude Opus has been lumbering along over the last week trying to turn what it did into a useful library. As far as I can tell, by trying to vibe together pieces of work fable did without understanding it properly. Opus makes mistakes constantly, misunderstands the spec, and it makes terrible engineering decisions. Earlier today it claimed it proved something was impossible. I asked it to think that through and it immediately backtracked and apologized. Was it right then, or is it right now? Claude has no clue.
I'm kicking myself for not using Fable more while I had it. Now that I've used fable, I'm not sure I even want opus any more. It might be more efficient in the medium term to just program everything myself until I have access to a similarly capable model.
I feel like Deepseek, opus, etc are only good at problems that have already been solved 100 times on github. They're like the iPhone 3G. Its exciting they exist at all. But subsequent versions make them seem like cheap junk.
I believe that local models are a necessary extension of the personal computer and I imagine that one could have had similar criticisms of early personal computers.
Maybe they've been more consistent but the strength of their position is not measured by my inconsistency.
Look at how much "trust me, I've got training, I know what's good, I know what's already right" is in their argument.
What is their actual point? That we can say across the board that good research must have easy-to-control experimentation and guarantee novelty?
Good research is field dependent; some fields are younger than others, some fields have an easier time controlling experiments than others.
I'm saying what matters is what people care about. My point about stances being political is because what gets funded is what people care about, not what can guarantee the highest confidence using research design.
My point is that their stance is political too, because it says 'I don't care about this like how they care about this, so I think it should get cut'.
Their position is not some innocent defense of empiricism, it's a political stance that says "these questions don't matter, I already know how the world works."
Rail lines in the US were not great examples of this. Many towns refused to grant right of way to the rail unless a stop was added which basically forced passengers to change trains. As a result, there’s were so many changes it took two to three days to get from say, Chicago to NYC when it should have taken no longer than a day
Interesting, didn't know about that Github button.
Id like to not depend that much on Github, and instead I'm thinking about having some kind of archival repo where noisy boring historical stuff can be saved (but I'd delete the branches in the main repo). Or renaming historical branches, eg adding an "old/" prefix.
I suspect this person didn't read the README or the `--help` at all.
For example, `--cpu-moe` makes it not offload MoE layers to the GPU, which drops performance about a quarter, but only keeps the dense and important layers on the GPU, allowing you to have MoE models bigger than VRAM almost for free, but also free up room in VRAM for more KV cache. It does nothing on CPU-only.
`--no-kv-offload` also does nothing here: it makes it not offload KV cache to VRAM... he doesn't have a GPU to offload to, and this is the default there.
Again, `-sm` is only for multi-GPU. No GPUs here.
`--mla-use` is for models that use Deepseek's Multi-Head Latent Attention. Gemma 4 is not one of them.
`--merge-up-gate-experts` reduces matrix math complexity around ffn_up and ffn_gate tensors; CPUs do not have tensor units and this is unlikely to actually help.
MTP is also never faster on CPU-only, and this is documented. ngram-mod, however, may help, which it doesn't look like he tried.
This whole screed also reads like it was written by AI.
For many years, I observed the San Francisco Caltrain DTX (Downtown Extension, recently rebranded "The Portal"). This is the most important transit missing link in Northern California that is expected to connect two of the highest ridership transit arteries in the Bay Area and eventually unlock single-seat rail transit between Sacramento, San Francisco, San Jose, and points south. DTX is a two-mile tunnel planned to connect the rail line terminus south of San Francisco downtown to Market Street, where the BART subway has the 4 highest ridership train stations in Northern California. The combined project (DTX and Transbay Terminal, the already built train station it's supposed to connect to) is about 15 years late and many billions of dollars over budget.
What struck me is a complete lack of urgency and accountability, combined with out-of-control meddling by politicians pursuing completely unrelated goals. The project spent several years in EIR and initial planning, which is to be expected. Then for over a decade, San Francisco's board of supervisors held the project hostage because they wanted to demolish a freeway south of where the actual project is, while bolting on an unrelated and unrealistic tunneling project (the "Pennsylvania Avenue alignment") and taking over the governance of the Caltrain board (Caltrain is the least dysfunctional transit system in the Bay Area, so the Caltrain board was not too keen on this proposal). Eventually, after wasting many years and tens (hundreds?) of millions of dollars, they agreed to stop holding the project hostage, restructure the board (TJPA), and re-hire staff to actually plan the tunnel.
I've seen multiple project managers/directors come and go, and countless community input meetings happen discussing completely hypothetical project concepts. The money set aside for the project from the original Transbay budget is long gone, and numerous funding opportunities have been passed by because the TJPA and its stakeholders were not ready to plan and submit a viable proposal in time.
Here are some things I would want to change going forward:
- Transit projects should be centrally planned by the state government (i.e. a regional subdivision of an agency similar to Caltrans) with structured opportunities for resident feedback and authority to override most input from local governments. This should include exemptions from CEQA and other review, and strong eminent domain powers.
- The Caltrans-like agency should have a dedicated source of state funding as well as a mandate to own and lease out land adjacent to transit stations as part of its funding. It should have a budget to retain project management staff who accumulate long-term experience and manage multiple projects. It should have the independent authority to issue bonds on behalf of the state and be required to publish construction efficiency and ridership statistics.
- Labor unions should be systematically prevented from influencing the course of planning, construction, and project execution. Unions meddle and cause many delays and project complications.
Unfortunately, even a structure like that is not a panacea. If you look at CHSRA, it actually has some of the features that I listed above. When CHSRA was first started, the planning process fell victim to meddling from state legislators (most famously the one who forced the route to go through Palmdale), followed by many wasted years fighting NIMBYs and doing useless planning. Ultimately, the only hope for a solution to this problem is to insulate the planners from political interference, set them up with independent funding, and reduce the veto powers that California grants to citizens and governments.
Ok, so you think your responsibility ends with voting? You build and fund a giant death machine and "welp my vote for death machine leader doesn't count so nothing else can be done"?.
"I'm selling an AI security product and want to establish my brand. I'll post several scare-mongering posts on my blog every week and people like solid_fuel will eat it up because it's what they want to hear."
How it’s price dumping if they give you the model to run free on your own hardware? Follow this logic, are not all OSS price dumping and should be blocked as well? I remember Steve Ballmer once called for that!
I’m pretty sure the digital lords like that proposal a lot. Not so much about the serfs themselves, though.
They're behind, but they're catching up pretty quickly. I wouldn't bet against Chinese companies dominating this market in the medium to long term.
Nvidia could easily end up being the Tesla vs Huawei as BYD.
U-6 is 8.1%, but spain uses a pretty comparable base employment rate (u-3 equivalent), so fair cop that US unemployment is easily less than half.
Relevant: https://en.wikipedia.org/wiki/Baumol_effect - drives a lot of effects in the US where automation paradoxically makes non-automated industries insanely expensive (though not the whole story for certain niches e.g. healthcare and education)
The check can not be removed if it becomes before the access because in this case the program has no UB.
A compiler in C is not allowed to move (time-travel) UB before any observable behavior. This was never allowed by in C (int contrast to C++), but the wording was not clear, which we fixed in C23.
If there is no check, you are right that the access itself is UB and there is no requirement to trap in ISO C, but this is something compilers explicitly supports.
(and might be required by POSIX, but I am not sure)
