metalspot

In the early 90s I had an elementary teacher who was an amiga nerd and had a couple for us to play with. Best class ever.

The concepts of industrial reserve capacity and using dual-use consumer goods to subsidize military production capacity are well known and widely practiced historically in the US. China adopted this strategy from the US, and the US conveniently forgot about it for a few decades in order to justify selling off the industrial base to China, but at least based on public documents like the published U.S. National Security Strategy, I would assess with high probability that this is explicitly recognized and being followed now.

AI is not fake and it does work, but what I am saying is that from a pure systemic analysis perspective, you can do the numbers, and even if AI was complete fugazi, the benefits you get from the electrical generation capacity, and the ability to fund it through private markets, which bypasses Congress, and locks in commercial contracts (often with foreign governments) which will be almost impossible politically to reverse, would still make it optimal from a strategic perspective. That is my calculation, and to the extent that it is correct, I would assume that the US Military's strategic planning apparatus would arrive at the same conclusion.

AI compute has some unique characteristics that make it especially useful for grid management. Moving consumer compute to the cloud means that the electrical use of that compute can be centrally managed. In an emergency, you can cut electrical use for consumer AI by 50% or more, because chips run more efficiently at lower power, and you can shift workloads onto quantized models, reduce resolution for video output, etc, to reduce compute, which leads to minor service degradation but not interruption. AI datacenters are also adding massive amounts of battery storage capacity, which is an additional grid buffer. For every GW in capacity added by hyperscalers that is creating a dispatchable reserve capacity of 50% under completely normal circumstances (hyperscalers do this internally to optimize their own costs) and then that number goes up depending on the scale and duration of the emergency.

Interesting. I am not familiar with model internals at this level because I have only been working at the application layer so far, but will definitely research this further. When you get the paper published would appreciate if you can drop a comment with the link so I can read it.

This also comes with significant capability reduction. deepseek-v4-flash is very good in the < 250K range, then degrades between 250-500K, and is practically unusable after 500K.

[edit]

This is my observation from using it without an specific context engineering to optimize for Deepseek's cache compression and sparse attention mechanisms. I am pretty sure that if you specifically structure your context to align to the cache compression boundaries you can significantly improve performance in the full 1M context, but there is not much reason to do this, because if you design your outer loop to work with shorter contexts that solution is portable and more efficient, so I haven't bothered with a optimizing for DS at this point.

Deepseek is still charging cached input at 1/10th the price of any competitor.

For an example of my real token usage for a day with DS: Input (Cache hit) 530,949,760, Input (Cache miss) 7,875,004, Output 1,389,685 - it is still 1/5th the price of Baidu (the cheapest) and 1/7th-1/10th the price of US hosts.

Also, Deepseek platform is not the same thing as Deepseek open weights. There is a major misconception that the existence of an open weights model means that it is the same thing as the proprietary platform offering, but that is definitely not the case.

You really have to look at energy/capita and how much energy is embedded in exports. The gross numbers are misleading. The US wasn't building new electrical generation capacity because it didn't need it and there was no market for it (caveats apply, but in a broad sense this is the major reason). Now that the market exists the question is how much can the US actually bring online and how rapidly, which is a real challenge after decades of degrowth politics used to justify slash and burn consumption of the industrial base.

AI is really all about electricity. AI could be completely fake and yield zero value whatsoever and the US would do exactly what it is doing now because the AI bubble is what creates the market for building new electrical generation capacity, which is needed for re-industrialization. Also why our friends in UK/Europe/China are so busy pushing anti-AI propaganda to try to undermine this.

I don't disagree but something less bad is better than something more bad if those are the only two options. I only said Duolingo was better than Youtube. Duolingo is nuts. I get my children's spam from them and it is this crazy emotionally abusive manipulative therapy talk nonsense trying to get them to use the app. I hope they are using AI to write that. LOL. I can't imagine the actual human who would do that for a job.

"broad anecdotal consensus" very scientific. i don't disagree, but is it better than youtube? tiktok? instagram? (cocaine? meth? fentanyl?)

Can you point to any serious studies that show the opposite? No. They don't exist. There cannot be any serious studies until there is a generational cohort that can be evaluated for long term impact.

But on a personal level I would vastly prefer my children use even a crappy AI chat app like Deepseek than watch Youtube. If you have young children and pay attention it would be almost impossible to come to a different conclusion.

literacy causes profound brain structure changes and psychological effects, and mass literacy results in profound social transformation. which is why people have be trying to prevent the spread of literacy for thousands of years, to this very day.

an LLM is a next token prediction function. of course it matters how you use it, and if you want to use it for an educational purpose, that has to be intentional, so i agree with you on that.

chatgpt and all the other ai powered scams and attention monetization apps are horrible, because they are just using improved technology to accelerate the exploitation of consumers (both adults and children) of media and app platforms.

the solution here is not to say the technology is bad but to use it to create competitive products that are actually beneficial.

code review is a dead end. the only way building with AI makes sense is in secure systems that can run unaudited code.

code review has always been a liability fig leaf. it is much harder to understand a system from reading the code than writing the code. if AI can write code 100X faster than humans it is simply impossible for humans to do real code review. effectively it is just pretending to do code review, and then running unaudited code without the proper systemic security guardrails in place.

Moving module by module is definitely the safer way to do it, but at the module level you do the same thing. You have to have it feature exact so that you can replay test against the same interface and do progressive roll out. The difference with AI is that the module size you can one-shot goes way up compared to traditional human rewrite.

same mess as before, just re-written

The likely path here is probably rewriting from node.js/ruby/python/etc to go/rust/c#/etc so a feature exact rewrite that passes all tests and can return identical responses for replayed production requests is not the same mess at all. You can do all sorts of refactoring, bug fixing, etc, while maintaining exact feature compatibility. The other major thing is back-filling exhaustive test coverage with AI, which then makes agentic coding much more accurate, because the feedback loop from failing tests provides the context for AI to self correct.

A test suite and code are two complimentary representations of the same logic, so using AI to grind test coverage to really high levels (90-95+) gives you two independent inference paths through a model, then the feedback signal from test failures gives you the mechanism to drive these two distinct generation paths to convergence.

Rewriting while changing features is the worst idea ever, which almost always leads to failure, and is the reason why "don't do rewrites" is a widespread rule.

Whether it is humans or AI, the correct way to do it is always a feature exact rewrite, so you can do comparative testing on both systems, and progressive rollout, and then you start adapting features.

An LLM can do this easily if you have a good test harness

How would you know? Are you an expert in the language and culture that its texts were originally written in? Even an expert would qualify their interpretation by saying that it is only an educated guess. Look at how easy it is for humans to misunderstand each other, even in direct verbal communication, then extend that to the level of misunderstanding in email and text messages. Now tell me how well you think modern people actually understand the meaning of something written thousands of years ago.

Every single SaaS app will be completely rewritten in the next couple years. Fair warning. If you aren't on that train you won't have a job.

Exactly. AI opens up a massive development frontier of projects that were simply impossible before. It does differentiate though. In the old world most "software engineering" work had nothing to do with software engineering so being highly skilled, educated, and experienced in software engineering made very little difference in compensation, position, promotion, etc. "software engineer" isn't a real thing anyway. coding is a secretarial job. you have to move beyond the mindset that coding is doing something useful. it's not. it is a means to an end and now better means exist. if you want to be an engineer you have to think in terms of systems engineering and building systems that deliver defined externally testable capabilities. in a few more years the idea of reading and writing source code will seem as absurd as reading and writing asm seems today.

this is a modernist misconception. ancient religions are philosophical frameworks and the texts are internally coherent arguments for those frameworks. They are just expressed in language that modern audiences without formal training don't understand.

priming obviously works. it is inherent in the physical structure of the brain. nerve endings from various all over the body terminate in different physical locations in the brain, and then these are interlinked in an adaptive biological mesh network. nerve responses are triggered by cascades of correlated impulses across different nerve endings, so at a physical level a strongly correlated signal from any two major nerve groups will trigger a sympathetic response when one of them fires independently. the experiments that social scientists come up with may be flawed, but this is definitely how my brain works, so i don't need a study to prove it.

No. Google and Apple are already rolling out the infrastructure. FCC is requiring IDs for phones. It will be opt-in for a while but the acceleration of AI risks will inevitably lead to an infrastructure level lock down of the entire internet. 99% of the population will accept it without hesitation because it is no different than what they do already. Everything you do online is already logged and tracked. The only thing moving to digital real id does is prevent criminal activity.

The entire internet will be real id verified soon. You won't even be able to send packets without a hardware digital signature linked to a real id. Might as well get used to it.

AI and remote interviewing have definitely exacerbated the problems that existed before though.

Leetcode was always weak, but now that it is easy to cheat on it is a negative selector, because the cheaters do best. Leetcode was originally supposed to be done in-person on a whiteboard to assess a candidate's collaborative problem solving skills, but with remote interviewing it has evolved into writing passing code with minimal or no feedback.

The real problem is that engineering departments are now filled with leetcode grinders and cheaters, who all live in permanent fear of being replaced by AI, and so any candidate who doesn't fit that paradigm is a threat that must be eliminated at all costs.

The chart at the top of the article makes it clear that the entire thing is pure fantasy

hallucination is good for tasks that have an external oracle like computer programming

The US is primarily powered by oil and natural gas, has massive domestic capacity, and locked in supply all over the Americas. China has a completely different energy mix and they move from a position of competitive disadvantage on ICE cars to competitive advantage on electric. Rapid electrification for China is all win, but for the US and our partners in the oil business, it would mean stranding trillions in capital investments that still have decades to run, so it just isn't going to happen.

But also a financial one

china's advantage is manufacturing capacity. giving away the model is a loss leader for the hardware business. they have a big gap on chip manufacturing. the only way to close that is by developing more efficient software. an open ecosystem is the best way to accelerate innovation of software development. since they are behind the US on model capabilities they aren't really losing anything by making the models open weights and being open about the performance enhancements techniques. but the open weights models are not necessarily what are running on the platforms or what they have internally. deepseek released v4 on their platform about a month before the open weights release, which I would guess was done to expose it to adversarial testing, so they could fine tune the removal of capabilities from the open weights model. (but i may very well be wrong)

I see it as a chance

absolutely. AI is the same as any other software, and open source has to integrate, adapt, and lead to make sure that open source values continue to propagate.

my personal approach is to focus on developing with open weights models, so that my work is optimized for them, and leads to their development. proprietary labs are free to copy, but they have a structural cost disadvantage. my objective is that open weights models remain competitive on capability but lead on capability/cost.

the user prompts and harness used for development are much more valuable for training than the final source code.

my approach to open source development with AI now is to include all of the agent sessions used in development in the repository, which makes this data freely available for training for both proprietary and open weights models, but that is just my own approach. every open source developer ultimately has to make their own judgement on the best way to integrate AI in accordance with their values.

How is this gloating?

he is emphasizing that they used their own words against them. everyone knows the security threat is a pretext. the message is that he is smart and they are stupid and he won, which is what I call gloating.

"Those trying to misdirect and tie this action to the prior DoW/Anthropic issues are wrong."

an obvious lie, which is inserted to emphasize that it is a lie. when you purposefully lie, not to deceive, but with the intent that the counter-party knows you are lying and must accept the lie, that is an assertion of power.

This is obviously political and the entire narrative is fabrication.

David Sacks is publicly gloating about it: https://x.com/DavidSacks/status/2065853007619588171

I can't really say that Anthropic didn't get what they deserved. They exploited security threats to sell their product and play political games, and now their rivals are rubbing it in their faces.