Yeah, 2000 years sounds like a lot till you do the math. Apparently astra was trained on 100k Blackwell gpu’s. So just over 7 days to crack 1028-bit rsa on that cluster…
Not to sound like a mark, I try to not get attached to any of these providers.
I've jumped between copilot, claude, gemini and chatgpt since the start of the year. chatgpt wasn't even worth looking at early this year.
Anthropic has the smarter models for sure, and seems to be default in corporate. However, the amount of budget you get with GPT as a user is much better, the harness feels more polished, and the models are faster. They are also much nicer to work with, I can just read the output for the most part. With claude I get pages of text and need to skim to find where the actual information i need to care about lies. So much more cognitive overhead.
Sol is smart enough for anything I've thrown at it, it's not one-shotting like fable, but I'm more willing to actually go back and forth with it, and it's likely producing better output to keep a human in the loop rather than trying to solve the world independently and making multiple incorrect assumptions.
I think GPT sees the market changing and is correctly repositioning themselves. Anthropic is down the wrong road, and if they don't correct course quickly I'm sure many of those enterprise contracts will start pivoting.
> and it's likely producing better output to keep a human in the loop rather than trying to solve the world independently and making multiple incorrect assumptions.
I was just discussing this with a co-worker yesterday. I would really like a model (or harness?) that worked with me instead of for me. Walk me through its choices and decisions, let me correct it and guide it along. I would be way more confident in it's output, I would be more familiar with the changes that are being made, and it would make reviewing the final code way easier since I was making the decisions along side it. I'm sure it would also reduce the "brainrot" we're all going to experience the more we hand work to these models.
does "Superpowers" exist outside of the Claude ecosystem? I feel like I've become hooked on this workflow - honestly I could take or leave the models ... it's the workflows and the way that they essentially create tightly focused loops over multiple sessions that I've becoming fairly dependent on.
Aside from the skills mentioned, I recently found that asking the model to create a simple HTML presentation to walk you through it's proposed design can do wonders. Fine tune your prompt to your liking (language style, what to include, what not, etc). Then, read that thing thoroughly, and keep asking questions and iterating on the plan until you fully understand what's about to be built, and are happy with it.
I'm ok with that. When the.major AI players implode under debt, it will create a massive proliferation of folks who go on to start new businesses unencumbered by the debt and bad decisions by the current leaders.
The collapse of the massive overvaluation and circular investments will be the undoing of several major tech investors and companies that have a serious contribution to the things in tech we don't like. It will be bad for the economy, but it will also weaken the abhorrent control that those companies have over regulation in the United States.
It's gonna be a rough time, but it's pretty clearly necessary.
The only big problem is that if that collapse happens under the current administration in the US, they will be utterly incompetent to respond to a real domestic crisis, since they have been virtually unable to do anything without creating more problems.
This is all great to get full unfettered access to your own system, as life should be.
I’m sure Xbox and PlayStation security groups are a little nervous right now though. Getting ring-0 on those machines is near impossible, but once you do then everything else becomes wide open
Not sure if it opens up that much on them as far as their security processors go. Modern consoles already treat DRAM as completely untrusted (an attacker could just sit on the DRAM bus and sniff/issue requests there).
The Xbox One for example encrypts all the DRAM it uses after it gets out of the main CPU die. See this part of Tony Chen's presentation https://youtu.be/U7VwtOrwceo?t=956
In the README it talks about accessing and modifying the code of the Platform Security Processor including accessing the keys, how is what is on an Xbox different?
The Xbox does not use AMD's Platform Security Processor, they have their own custom controller on die with Microsoft hardware and their own custom BootROM. This custom security processor actually has its own bit of RAM entirely in the die and doesn't even go out to the DRAM at all.
But the million dollar question is, do they rotate the IV per block or do they opt for speed as AMD and Intel both did by deriving it from the address and thus are vulnerable to someone sniffing the bus? (Note that older Intel server hardware wasn't vulnerable in this way but tracking IVs slowed everything down so they did away with it.)
wouldn't request after DRAM controller be unencrypted ? Would need to have different key per memory type and even then you could do some damage as realistically it won't have number of keys equal to running processes
I hope in my lifetime we discover a way to predict earthquakes better, the danger to life from them would be significantly less if there was sufficient notification.
We can alert for storms, tornadoes, volcanoes, hurricanes, typhoons, and even somewhat extraterrestrial disasters like asteroids.
The earth itself must move in predictable patterns similar to the weather, and I'm sure there is decades of research around this I'll never even approach to understanding. Just surprising to have such a big knowledge gap that affects humanity year round
We are actually pretty good at predicting the long term for specific regions. This allows building codes to be based on the seismic activity of the region. Sadly this isn't something that's practical in many parts of the world
Really I suspect that's the only hard part. If we had accurate 3d models of the structure (well 4d, 3d models over time) it would probably be reasonably easy to estimate how much strain it can take, and how much strain there is... Maybe not predict-to-the-second levels of detail but at least to predict that a quake we see is likely to be a precursor to a much larger one or is likely that peak.
But getting that data seems... practically impossible at our current tech level.
5% within a week is a very high set of odds for such rare events. Or to look at it from the other more relevant direction 70% of magnitude 7 or larger quakes have detected foreshocks...
In fairness, games didn’t start employing encryption on the content till the 7th generation. So unless you’re applying ips patches to a 360 game you’re not breaking any DRM
Playing the title you backed up and patched is a different matter entirely
Courts have ruled many things besides encryption count as anti-copying measures, and that the laws apply retroactively, so it’s not so clear such patches even on 360 are legal. Try selling such patches and test it.
It will be a push and pull and I'm curious where we'll be in a couple of years.
I can fly at 100mph if I let AI run loose and with a bit of steering I can get it to output what I'm looking for and generally pass verification and tests.
If I care about the code though, and I want to keep it maintainable, the amount of time and tokens I need to spend correcting and iterating on the output quickly eats through much of the initial time I saved, to the point where I'm unsure if I'm actually saving much time at the end of the process.
With hobby projects I lean on quality more, and the async nature of AI also makes this much easier to make progress without needing my full attention to do so.
In the corporate world, there's both the pressure to accelerate with AI, but also maintain code and product quality. The dials of one way or the other are more obvious now, but I don't believe it's possible to do both with the current models and harnesses without exponential cost.
What is interesting though, is I'm now leaning towards faster models rather than smarter ones.
Intelligence lets me bite off larger chunks of work at once, and trust the model to behave without having to watch it intensely, but doesn't seem to drive down the number of iterations required to hit my desired quality.
Faster models means the iterations I'll have to go through regardless will complete much faster and gets me closer to a proper flow state. Models will keep improving, but maybe we're getting near "smart enough" and the race will pivot to performance > intelligence.
I also spend a lot of time at work trying to get the code into a shape I deem acceptable to open a MR for the team, without looking incompetent.
At home I stopped reviewing the code for my mobile app, and I do not feel this has affected maintainability.
With 150k LOC, the AI still performs changes as easily as it did three months ago, when I started the project.
If anything, the new models made it easier.
When I read all those posts from people advocating a very hands-on, understand each line approach, I speculate many come from a position like me at work or have never seriously tried 'vibecoding' a large project from scratch with SOTA models.
People routinely post that things I work on for 30 hours/week are impossible.
150k sounds like a LOT… what’s your use case for the app?
It seems for any particular problem/comexity there is an estimate of how many LoC is needed (written by humans, good coding practices). And after reaching maturity (and not increasing scope) the LoC converges.
My experience with LLMs so far is it’s a constant battle to not explode the LoC. There’s always pressure away from “maintainability” and a “small elegant cohesive” codebase.
It is in media management. I cannot talk about the features in detail, but I would not be surprised if it can be done with half the line count.
What I was trying to say is that maintainability appears largely unaffected.
Doing manual QA testing and iterating on improvements still works great.
The question then is in which cases "small elegant cohesive" is worth it, when the implementation effort increases tenfold. For the product I am working on at home for 30h/week, I can confidently say it is not worth it.
Clearly hyperscalers are seeing datacenter commitment exceeding demand, and are now running PR campaigns to have the public push back against construction to get them out of their contracts.
Joking conspiracy theories aside, perhaps the over-commitment was part of the plan with the expectation that there will be rejections from government
Indeed a classic (upvoted!), but it presupposes that there is a theory of everything that can be used to perfectly simulate the universe, and I'd like to remind everybody that we still don't know whether such a thing exists. Nor do we know whether it's possible to simulate "just" neurons and get meaningful results on some higher level. People have been trying at that for decades, so far without success. (nice overview here: https://ccli.substack.com/p/the-biggest-mystery-in-neuroscie...)
Thank you for the link to the post. It was a fascinating read and also entertaining.
After seeing the story of the fruit fly brain simulation I figured stuff like the nematode would be trivial.
The lack of active plasticity in LLMs causes a lot of frustration imo. Having agents make the same mistakes over and over and not be able to correct course without prompt engineering is the biggest symptom.
Obviously a static mind is more helpful from a performance perspective, especially if you consider model deployments into gateware.
Maybe this will be an inherent limitation to LLMs that won’t be solvable without a new approach.
I was curious what fruit fly brain simulation you meant, and after fifteen minutes of looking around I ended up at a LessWrong writeup by an author from something called the Brain Preservation Foundation -- which presumably has an optimistic view of the field, but its director is an author on the original fly connectome paper, so this should not be utter crankery: https://www.lesswrong.com/posts/ybwcxBRrsKavJB9Wz/no-we-have...
It seems that there was a perfectly legitimate simulation based on the connectome that got published in 2024, and a very bullshit-y startup that briefly went viral later when it claimed to upload a fly to a computer. Not sure which one you saw.
> [...] but it presupposes that there is a theory of everything that can be used to perfectly simulate the universe, [...]
LLMs are nice here: we don't need a theory of everything to simulate them. You can indeed use pebbles on the beach.
Scott Aaronson said that in theory a finite look-up table can win eg a ten minute Turing test (any finite length is fine). Of course, to borrow his argument further, the naive table would be galactic in size. So the real magic is in being able to compress the responses of that table into something human sized / LLM-server sized.
> but it presupposes that there is a theory of everything that can be used to perfectly simulate the universe, and I'd like to remind everybody that we still don't know whether such a thing exists
Is it even possible that a complete simulation of the rules of this universe (let alone the content of it) could exist in this universe?
I am no expert but I did manage to read most of GEB [0] so I feel fully qualified to say that it feels like something Gödel would have opinions about?
[0] the honest nerd's "did sleep in a Holiday Inn last night"
As an analogy: you can have a zip archive that expands to itself. Or even more so: to itself (the simulation of itself in itself) plus something arbitrary (standing in for the rest of the universe).
It is entirely possible (if perhaps unlikely) that the correct theory of everything is a relatively straightforward expansion of the current standard model.
I would argue that for the situation of the comic to occur, it's only necessary to have a theory of everything that's sufficiently good to simulate a universe that's close enough to ours that it's hard to notice the difference.
Our bodies could just be rocks on a beach, but whatever it is about us that is capable of observing the output of our own internal thinking (preventing us from just being p-zombies) is not contained within the beach.
Well, thermodynamics is considered pretty fundamental, and the second law of thermodynamics is pretty much all about information. (But your link already touches that.)
People also interpret the 'unitarity' of quantum mechanics to mean that information is preserved.
I think you're being downvoted because you're announcing it as an absolute truth, of which we do not have evidence.
However I am open to your belief (and not on religious or theistic grounds which I assume many people might think you're talking about) because people have (non-religious) experiences that lend to such possibilities.
Note the README in the Unsloth list of files: llama.cpp is working on a PR to support the gemma4 drafters: https://github.com/ggml-org/llama.cpp/pull/23398. Also note the PR submitter didn't experience much speedup with 26B (seems typical that MoE models don't generally benefit from MTP).
I do have the Qwen 3.6 (35B) MTP implementation running (in LM Studio; it doesn't need a separate drafter), along with non-MTP Gemma 4 26B, and I can see that Unsloth Studio can run the new QAT, but I can't see how you can run the assistant/drafter. Yet.
It's just a constantly changing landscape. Don't get me wrong, it's fascinating and for various reasons I am pleased I can keep up even slightly, but eeeehhh :-)
Yeah — that is the base QAT model, and there are safetensors weights for the QAT version of the MTP drafter, but there are no MLX/GGUF versions. I think the answer is a combination of:
1) Gemma 4 MTP is too fresh for off-the-shelf software to use anyway
2) "you can convert them yourself" which is fine, obvs
reply