This factory was built by the US government for WW2. It was operated by the government until the 60s where it was ultimately shut down. In 1998 it was later converted to a contractor operated facility and has been ever since.
We gave those companies so much money and they decided it was fine cutting corners and improperly handling the explosive ordinances under their care. We've literally given this company 4+ billion dollars, most of that in recent years with this year's spend being 270M https://www.usaspending.gov/recipient/1286c0e7-b6b5-8c72-290...
And for all that money dumped on them, what we've ended up with is the US government taking ownership of them mishandling large amounts of dangerous chemicals.
When someone talks about "waste, fraud, and abuse" This is it. This is what the $1.5T military budget is buying. Huge investments into an army of negligent private companies. No doubt they lobby well.
At this point, every hiring manager has seen every way a candidate can use AI to lie about their abilities and accomplishments, so it won't work anyway.
It turns out that people do still appreciate honesty. OP thinks that taking two years off for care-giving is a red flag to employers. I assure you it is not!
- Quoted phrase in the opening sentence
- Opening quote: “you need to lie”
To expand on that: you need to tell potential employers about your AI-powered SaaS or whatever startup that you've been working on for the past two years, how much you accomplished just by yourself with the help AI agents, how much you learned about using AI effectively, etc. That you just couldn't get enough traction in sales and funding and the area you addressed is now being broadly up-ended by AI, so you are looking for a more regular job
On the side you obviously need a demo of what you've built over these two years, but Claude can help with both brainstorming what your startup actually did and building it retroactively. In the process it will force you to actually start using AI for real which will be a nice bootstrap for your eventual real job.
Let me give you some harsh feedback.
You can't have "learned the AI" and still think coding in C#/.NET/AWS is a valuable skill. It's not; the AI codes better than you. It also codes faster than you can read it.
My strategy has been focusing on orchestrating highly automated workflows that can deliver software at breakneck speeds. I only focus on the business domains, security, and software architecture.
Try this to start: check out the /wayfinder skill from Matt and build something you could never have in the past. Extra points if it's something you can sell to a small business. A system for your local gym, for instance. Give yourself a single week.
Coding is not a job anymore. It has evolved.
The buzzword everyone is looking for is "high-agency."
No, not those agents, but yes, those will be essential.
Basically employers want someone who will start something from scratch and take it to the finish line all by themselves. So my advice, which others have echoed here: Build something useful -- end-to-end -- from scratch -- with AI -– on your own -- that you can show off as soon as possible.
The key point is, "high agency" is by definition not something you can just put on your resume; It is something you show, not tell.
Yes, you need to do this even as you go through the absolute hell that is a job search. But trust me, this will a) help get a much better job, and b) help in the long run throughout and beyond your career. This will be the most valuable skill in the near future.
A few other thoughts:
0. Get a $100 Pro subscription. Either GPT or Claude, doesn't matter much, but you do want access to Fable or Astra. The volume and variety of things you can get done in that quota is astounding. Codex offers a free quota with Astra if you want to get a sense of things.
1. Target a domain you are already very familiar with. You can draw on your long experience as an engineer, or any personal interests you may have. Say a Math tutoring web app that your students could use. This will sharply cut down the time to MVP.
2. Being an end-to-end thing, this will necessarily be something multi-disciplinary. Think of it as doing a one-person startup, so you're going to be forced to do everythig: design, frontend, backend, infra, and maybe even GTM stuff like marketing. Yes, you may not know most of these things, and obviously won't be great at them, but with AI, everything will be good enough.
3. This conveniently also helps explain the gap in your resume. Nobody needs to know you did this in a few months vs 2 years unless they ask.
4. Even when using AI, keep your hands-on skill warm. Most companies still do old-school leetcode interviews.
5. As an extreme backup this might even become something you monetize yourself.
Hope this helps, and all the best!
- Ordinary term used more than once in the same comment
- “stuck” appears 2 times
I run the oldest turbulence forecasting website on the internet. What I don't like about narrow-body aircraft on transatlantic flights is that you have all these big jets that can go up to 37,000-40,000 right away and fly over turbulence. If you're on on a A321XLR, not only is it a small plane, but you're "stuck" at 32,000 or 34,000 bouncing around where you could overfly turbulence if you were in something bigger and more performant.
I get the concern, but isn't that a bit melodramatic? For employment, nothing has changed: you get a job by being smart (and likeable, and dependable). To the extent that the job may truly require the use of gen AI - which is a small fraction of white-collar jobs - your employer will provide you with access. Just like they always did, most jobs provide you with the equipment you need, from computers to jackhammers.
It's more of a concern for businesses, because OpenAI or Anthropic or Google may now be the arbiters of what types of businesses are acceptable and what types are not. But... Big Tech already does that, it predates AI. If you're the wrong kind of a business, you can't do online payments, can't be on Facebook or YouTube, etc. Talk to anyone running a dispensary or a gun shop, it's a barrel of fun.
Further, mathematics is the one field where the practitioners have the power to say "no" and define their own rules of the game. They probably won't, but that's on them. In contrast to software engineering, there's no one telling them they have to use AI from now on.
- Quoted phrase in the opening sentence
- Opening quote: “They”
And this appears to be his personal math blog he was using to write his last book.
I imagine that he considers it a safe space to use such an acronym.
abelian group => a group which commutes, eg a.b <=> b.a
Schwartz-Bruhat functions => functions which pretty quickly go to 0 as their vectors go to infinity, so they kinda are "bumpy" in some finite place, then everywhere else is 0-ish
The Alpoge proof, if true, will certainly push mathematics forward a great deal. In the announcement, he lays out the general approach and says that he had 10+ "WOW" moments of insights while working with the AI on it. I have no doubt that there will be tons of follow-on papers clarifying and expanding many of the ideas that the AI explored. This will likely be the case even if the proof is wrong.
- Quoted phrase in the opening sentence
- Opening quote: “right”
But, AI muddles this a lot given the complete lack of provenance from the companies about where their data comes from and secrecy on the methods and transcripts. So far almost every major result from them has had this issue, and in the absence of increasing transparency from companies, this just pushes me more and more towards believing that the companies are trying to cover up bad behavior in the pursuit of these announcements.
- Quoted phrase in the opening sentence
- Opening quote: “GRH”
[1] https://en.wikipedia.org/wiki/Generalized_Riemann_hypothesis
edit: to all the people responding that LLMs are good at math, I'm going to paste a reply I wrote downthread...
I'm not arguing that LLMs aren't useful tools for mathematicians, I'm questioning the quality of discourse that amounts to "I pasted this into the LLM and it said there are errors"
- Quoted phrase in the opening sentence
- Opening quote: “right”
Or, alternatively, did people start working harder in some fashion, even on creative problems that require deep insights, because they felt it was their last chance?
Or, alternatively, are people just (sensibly) starting to post all their work-in-progress results to establish priority (on either the results or approaches used)?
I think we’re going to start seeing a lot more of these types of advances, especially on smaller models
I’ve been able to graft reasoning and math problem solving into a small, locally trained model with 6.5M params. And I’ve also been able to add abilities or improve them by “growing” the model to 8.5M params (this usually means redistributing the weights in a slightly different network size, then finetune the new size)
The model is so small that I can sometimes run dozens of experiments per day. There are a lot of papers and new models coming out. A lot of the stuff they are doing at 100B-1T still works at smaller scale (usually not the scaling inference-related optimizations, but rather things like weight programming, or a “looped” architecture). It’s a lot of fun to run at home on my laptop
I imagine bigger labs with bigger budgets (more GPUs) are mostly focused on 20B-1T+ models. But there’s a lot to explore in local “home models”
Also, if anyone from the team is reading, if you have access for researchers I'd be interested in doing some higher volume tests & harness tuning than I have hardware for personally.
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -2670,10 +2670,12 @@ ggml_tensor \* llm_graph_context::build_moe_ffn(
const float beta = hparams.situ_beta;
const float lb = hparams.situ_linear_beta;
- ggml_tensor \* act = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, cur, 1.0f/beta)), beta);
+ // gate/up are strided views of the fused gate_up output; ggml-vulkan's SCALE
+ // needs a contiguous src, so without the cont both scales fall back to the CPU
+ ggml_tensor \* act = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, ggml_cont(ctx0, cur), 1.0f/beta)), beta);
act = ggml_mul(ctx0, act, ggml_sigmoid(ctx0, cur));
if (lb > 0.0f) {
- up = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, up, 1.0f/lb)), lb);
+ up = ggml_scale(ctx0, ggml_tanh(ctx0, ggml_scale(ctx0, ggml_cont(ctx0, up), 1.0f/lb)), lb);
}
cur = ggml_mul(ctx0, act, up);
cb(cur, "ffn_moe_situ", il);- Quoted phrase in the opening sentence
- Opening quote: “extra high”
> While looping improves the model’s reasoning, it doesn’t enable its weights to store more facts—knowledge is mostly a function of total parameter count. Inkling 2 Mini’s score on AA-Omniscience, a measure of factuality, is in line with other models of the same size [...] but knowledge-intensive benchmarks doesn’t improve with more passes...
> Inkling 2 Mini overcomes the knowledge gap by efficiently using search and tools.
I tested the Unsloth Q4 version.
My prompt asked it to write a singleton class, single file plugin for WordPress that hooks user login and stores last login time as metadata. It has some code style requirements. The prompt also asked it to make no tool calls so it didn't search and only used its knowledge.
The model burned through 19200 thinking tokens on this task (which was the hard limit I had set) and didn't finish. Its thinking suggests it was going to make at least one mistake.
Both Qwen 3.6 35B and Gemma 4 26B can cope with this task, from memory, in far fewer tokens (even taking into account Qwen's tendency to loop in thinking).
Rerunning the same prompt with tool use allowed, it searched with browser use, pretty effectively, and got to a solution, but it still made the mistake the thinking traces suggested it would earlier (wrong hook name). Other than that it wrote good code.
But 20K-ish tokens in, I could then not get it to search to make the required correction — it would repeatedly EOT rather than actually run the tools it was saying it would call. Whether this is the model failing or Unsloth's harness and the model disagreeing, I don't know. Quitting and reloading the model to continue the conversation didn't help.
It worked through my usual MySQL-5.0-compatible max score per category prompt nicely, though it didn't volunteer the suggestion of adding an index for speed, but it doesn't outperform LFM 2.5-24B-A2B, which is a non-looping model of similar size. It does work through SQL in its thinking traces quite nicely, though.
Car wash problem got a correct answer quite quickly, albeit with a surreal caveat (why do they always seem to need to offer a caveat, no matter how weird)
So far, I think I would run Muse Glimmer or Nemotron Bolt over this; Muse Glimmer would output tokens more slowly but still get to a result much more quickly.
- Quoted phrase in the opening sentence
- Opening quote: “We're talking about tens of millions in artificial intelligence costs.”
TIL that 1 million Icelandic Krona is just $8K USD.
If you are looking to escape the "slopfest" and to do high quality work made to last decades, you are in for a rude awakening. For most of my life, I was subjected to overhearing frantic and frustrated phonecalls where my stepfather had to deal with shoddy work, lazy coworkers, and general bad behavior. This was true when he was a journeyman and when he was a master, foreman or not.
I used to think it was a feature of that line of work, but when I started working in software I just realized it was the nature of any "builder" role; you the four axes of people who care, people who don't care, people who want to get shit done, and people who just want to get paid. You and your colleagues are all going to chart and then be thrown against projects that demand some bar of execution on some timeline, with all the lovely conflict and shortcuts that implies.
Compare Elliot Glazer a month-ish ago:
> I occasionally talk to experts on all the MP's. Their rough summaries on what is needed to solve them usually go:
> Hodge: not a clue
> BSD: not a clue
> RH: not a clue
> P vs NP: not a clue
> Yang-Mills: "Well..."
https://news.ycombinator.com/item?id=50252673
So if we're in "not a clue" territory, maybe that's that.
I fear that it gets muddled again with e.g. the concurrent human announcement and that in retrospect RH turns out to be easier than expected, like many other problems before.
Math is hard, but how hard math is is also hard...
It seems we’re making tremendous progress on the back of existing abstractions and concepts with the brute force ability to generate language and verify proofs.
For example, Tao says about the Riemann proof:
> appears to use the full power of modern developments on the Langlands program and arithmetic geometry
What I wonder: how do we create new abstractions and concepts? Like whatever “Langlands program” is? That then feedback into the system, and allow for even more sophisticated mathematics.
Are we assuming AI will do this? Or does this still need to be something at least agreed-upon by a mathematics community - in the same way programmers talk about a “linked list” without having an hour long conversation to define it axiomatically? And without that social aspect, without breaking down long proofs into Lego block common abstractions, math may plateau? In part because we lose the ability to ask meaningful questions?
For inexperienced engineers it's that "your job is figuring shit out, not typing the code in", which for some people is like telling them that their dream job doesn't exist/the mental model most of their peers operate under isn't true.
For a manager it's that you can't really assume that everybody has the same personality/relationship to work/experience as you do, and that you might have even forgotten what it was like to be in their position, or not noticed how different some of your colleagues were from you. Which is also like telling them, "it turns out the best manager for your team is not the one you would want for yourself, and this job is going to be very different from what you thought".
I think it's very important to have a large enough sample of non-reporting colleagues across the range of experience. It helps you passively form a model of what normal career and even just life development looks like, and when you're in situations like this where you're not sure whether you're the asshole, or the other person is, or just confused, you can kind of in retrospect see how the near-peer may have navigated the same thing.
Same goes for interacting with enough non-developers and business people/customers/higherups to get some about the purpose of your role. Your role is downstream of what you do for them.
Yes, that is how most functional organizations operate.
Architects are a bad meme, along the same lines as scrum masters and product managers. Good software engineers spend more time on design than implementation, and the best are helping their teammates with design, and auditing designs of adjacent teams to ensure integration success. So architecture is an important part of the job, but that doesn't mean that you can automatically declare a role with the title "architect" and make the whole organization better at design.
Architect roles are a bad meme because the architect always ends up knowing less about the system and how it is used than the engineers and support staff, yet they have more influence over the decisions affecting those things. That's the essence of an organizational anti-pattern/bad-meme any time it shows up: influence over resources anti-correlated with decision quality. Additionally, architects often report to non-technical folks fairly high up the chain. Those people are bad at screening for such a technical role, sometimes even making the mistake of replacing an architect with an external hire.
After getting complaints, I started getting narrower and narrower in my ticket writing, until I realized many newer software engineers literally want tickets like "modify function x_y_z() in python_file.py so that it also returns..."
Try that with photography, painting, or music: "AI music has promoted musicians to architects".
- Quoted phrase in the opening sentence
- Opening quote: “God says”
Scientific American, May 19, 2026.
https://www.scientificamerican.com/article/the-riemann-hypot...
- Quoted phrase in the opening sentence
- Opening quote: “paradise”
I sincerely don't understand how on Earth people make the leap from our current trajectory to "paradise". What makes you think any of those things follow naturally from where we are today and where we're going?
Every time someone says 'current AI is not enough, we need new architectures', it turns out that current AI was enough and all we really needed was more scaling.
I suspect that these "smart, technical people" you speak of have more nuanced views on AI than you realize. Myself, for example, a self-styled smartypants:
Like the dot-com bubble, I suspect that certain parts of the current AI landscape are unsustainable, and I suspect/hope that those specific features of AI will crash and burn, while other useful bits remain.
AI usage like we see in this post here--AI that assists with discoveries in math, medicine, genetics, astronomy--I think they're great, and I hope they will continue to advance human knowledge for the foreseeable future. They assist experts and professionals rather than replacing them.
In contrast, generative AI ("creative" AI writing, illustrating, music, etc) is a pox on civilization. It is industrialized low-effort plagiarism that harms artists' livelihoods, discourages new art, makes everything bland, and pollutes the world in multiple senses of the word. I hope it dies from a combination of negative sentiments, and too much real-world cost once token subsidies go away.
Then there are the AIs that do things like voice reproduction. They have some arguable positive uses (e.g., assistive technologies), but many negative ones (e.g., phone scams, replacing human narrators with copies of their own voices). I think we're better off without these apart from niche accessibility uses. I think it should be legal peril to reproduce a voice without permission.
Apart from helping with some cinema CGI, deepfakes and deepnudes seem like a net negative for civilization.
Knowledge-based AI like Google's AI search summaries might just have built-in expirations. I think people will burn out on being vigilant about hallucinations. And when the bulk of their training data is made by AIs, it'll be hard to prevent model collapse. The cost of human intervention might make it too expensive for the wide use we see today.
Coding AIs are a grayer area. I think it's highly problematic that they are basically open source plagiarism machines, but I begrudgingly believe that society will decide that the upsides outweigh the downsides in the long term. But I choose not to use AI in my own coding projects, and respect others who do the same.
To: Buckmaster
From: A friend at OpenAI
Subject: Openai is about to do something really stupid.
"I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, 'Why would you ruin your career?' I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, 'If you don't want me to be nice, then I don't have to be nice.'" (Sébastien Bubeck later apologized on X for the remark.)
"Sebastien twice asserted that he wanted Levent removed from authorship..."
"The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAI model had resolved it.""I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer." (OpenAI subsequently stated publicly that while unlikely, it could not rule out that de-identified data from their product usage improved the model.)
- Quoted phrase in the opening sentence
- Opening quote: “if the exponential time hypothesis is true,”
People were talking about superintelligence in 2021, but "by the end of the decade", not "next year".
- Ordinary term used more than once in the same comment
- “*.lean” appears 3 times
mkdir -p riemann_proof && cd riemann_proof
# fetch parts 01–14
for i in {01..14}; do
curl -fSL -o lean.tar.gz.part$i https://alpo.ge/lean.tar.gz.part$i || { echo "failed: part$i"; break; }
done
# join and decompress (creates ./lean)
cat lean.tar.gz.part{01..14} > lean.tar.gz
tar xzf lean.tar.gz
# size, file count, line counts
du -sh lean | cut -f1 | xargs printf '%s on disk (lean folder)\n'
find lean -name '*.lean' | wc -l | xargs printf '%s .lean files\n'
find lean -name '*.lean' -print0 | xargs -0 cat | wc -l | xargs printf '%s lines total\n'
find lean -name '*.lean' -print0 | xargs -0 cat | grep -cv '^[[:space:]]*$' | xargs printf '%s non-blank lines\n'- Quoted phrase in the opening sentence
- Opening quote: “months or a year”
No one said that. If you look at e.g. the Metaculus predictions, you can see AGI has been hovering around 2027-2030 for 5 years.
https://www.metaculus.com/questions/3479/date-weakly-general...
https://www.metaculus.com/questions/5121/date-of-artificial-...
If you do want to see something done about it and this isn't just another prediction or a vague warning, more power to you, but otherwise, I'll wait and see
- Quoted phrase in the opening sentence
- Opening quote: “we got reimann hypothesis before GTA6”
In the ~150 years since it was posed, it has become a lynchpin for hundreds of downstream results that assume it is true. This proof would immediately confirm all of those results.
More practically, there are also some algorithms that will now have guaranteed theoretical bounds and it will likely eventually lead to some more interesting algorithms in the future.
It's kinda the math equivalent of "we found life on another planet".
I’m not claiming to be “top human intelligence” but I’ve published in premier venues in my field and I can say confidently that none of my human colleagues can keep up in conversations on my research niche like Claude can.
I'm just not convinced that we're swapping one world for the other. We're getting a hint that will probably help us enormously with what was always the main objective: understanding the result.
The worry I have right now is this though:
> The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail, in my view of my path I have just taken a step.
this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
And we're back in the world of 'magic' and 'belief'. Yes, there are formal ways to prove that this is correct, but if the proof is so convoluted and hard to read, is it actually something a human will ever find the time to go through and understand? We're going to start delegating (oh well, we're already doing that, I suppose) the understanding to AI, so it can just help us condense it down to smaller pieces that are digestible.
Still, very strange dizzy feeling, much more than Navier-Stokes
- Quoted phrase in the opening sentence
- Opening quote: “slopstrap off”
But the name "Rebel Theorem 5.0 ML/AI model" makes me suspect they are full of shit.
I also went through 3 hinges on my 1st gen Framework 13 laptop, because the hinges were so weak they couldn't hold up the screen under its own weight.
My thread cataloguing my issues with the Framework 13 got so long, the moderators closed it.
My trackpad stopped detecting clicks. So did the replacement that I bought because it was out of warranty.
One of the buttons for the expansion card removal fell off. Framework replaced that under warranty.
The speakers would cause the keyboard to rattle until I took off the shift keys and adjusted the stabilizers.
I really wanted to believe in and support Framework. I support their mission. But man, everything I've gotten from them has been junk.
After it got to the front page of HN, Framework promised action when asked by Ars Technica. Three months later, people are still getting their laptops bricked and writing about it on their forums, and the promised fix to add BIOS flash recovery and making bricking impossible still hasn't been released...
People have been complaining about Framework's support practices for years and even their investor (Linus Sebastian) has done "secret shopper" on their laptops and run into issues, and the response has always been "we'll do better." I am just not seeing the improvements.
I want there to be a good repairable laptop, but the company has to follow through on their promises and treat their customers right.
This should be a case of the engineering team getting the faulty device for debug (and ideally getting all the relevant information so they can do the work with some semblance of parity with the customer's setup) -- and the other half is the customer gets a functional piece of hardware to swap in.
If Framework is worried about sending out hardware before receiving hardware, the policy is a simple one; when we get the hardware, we'll ship your replacement.
---
and the amount of back and forth for the serial number of part X, then Y, then firmware of Z, etc.. maybe there's a technical reason for that but i just considered myself beat into submission and gave up.
This is the standard 'debug over the phone' thing and engineers will do this while collecting information to try and recreate the problem you're having. It may be because the engineers are too close to the process? If you think about it, the RMA process is actually a sales process. You're essentially selling an already sold user on continuing to be a user.Me and my GF filed a report with every proof to mastercard for a chargeback (that framework contested), but after Mastercard investigation, they ruled in our favor.
This is a scam company. Never ever buying again.
I bought a chasis kit recently and the touchpad is faulty, or acting faulty ( physically getting stuck depressed), and the amount of back and forth for the serial number of part X, then Y, then firmware of Z, etc.. maybe there's a technical reason for that but i just considered myself beat into submission and gave up.
after mainboard upgrades, touchscreen upgrades, and various other upgrades over the years and being excited about framework, it's weird to have all of my enthusiasm get derailed over support for a touchpad.
- Quoted phrase in the opening sentence
- Opening quote: “blow up”
I wonder if we have enough compute, and/or modelling accuracy, to do a molecule level fluid simulation to see what would REALLY happen as this blow-up scenario is approached? It's not necessarily going to be weird.
I don't, at all. In fact, and this may be controversial, I kinda like the OpenAI approach of crediting the model, and not an author. The model did the vast majority of the work.
I feel bad for folks like Kevin Buzzard, doing the FLT Lean formalization by hand on a grant and then watching AI blow past it.
I also feel bad for the young people entering the field in such "interesting" times.
This is mostly OpenAI and Anthropic racing each other with a nearly unlimited token budget, hard to feel sympathy for that.
- Quoted phrase in the opening sentence
- Opening quote: “sleep”
https://research.google/blog/the-sleep-framework-a-new-way-f...
> During the offline sleep phase, the model unlocks pre-allocated Mixture-of-Experts (MoE) memory blocks and updates its numerical weights using self-generated practice examples.This is the best I came with so far, but I have to admit how slow my main projects move, if at all.
But that makes LLM speedup like 1.5x the manual-typing coding speed, vs 50-100x you can get if you vibe all the way.
I still use LLMs for troubleshooting, searching, codebase knowledge gathering, solution exploration, writing tests and documentation and whole-codebase changes and there the speedup is dramatic. However I don't feel I become a better developer for doing these kind of tasks without LLMs.
For documentation and tests though I feel I am slipping on reviewing the output because it is just so boring that I often "surrender" the task to the LLM which generates bloated tests and bloated docs.
- Quoted phrase in the opening sentence
- Opening quote: “anthropic tax”
The gigawatt-sized asterisk on all of this is that AI companies have to capture that $1 billion. Anthropic has to ten-fold increase it's revenues or it goes bankrupt under it's current commitments. It has to increase it's revenues another 10 times to reach the $10,000,000,000,000/yr revenue required to support it's current valuation.
Anthropic put their Total Addressable Market at 35% of Global GDP (~$40 trillion/yr). They're planning to capture all of the money.
The second gigawatt-sized asterisk is that because of the fine handiwork of Mr. Huang, China has made very competitive AI models available for free, so the actual ability for AI companies to capture the value of the labour they displace ends up being quite limited. Thus a problem in the predictions: Something as to break, but it is as of yet unclear what that will be. Maybe Chinese AI will be banned, maybe Anthropic & OpenAI go bankrupt.
Any place-value numeral is a shorthand for writing a particular geometric series. Only considering numbers between 0 and 1:
0.abc… = a*10^-1 + b*10^-2 + c*10^-3 + …
If there exists one last nonzero digit in the place-value representation, there is one last nonzero term in the series, and the sum is trivial to compute.But if the place-value string is infinite? Well, we know that infinite series can have finite sums, and in particular all of these place-value series converge and have finite sums [exercise: show why].
Specifically, the sum of a converging infinite series, and thus the value of the corresponding place-value expression, is defined as the limit of its finite prefix sums. The general formula for the sum of a converging geometric series such as 0.9… is
S = a / (1 - r)
where a is the first term and r the ratio of two consecutive
terms. In this case S = 0.9…, a = 0.9 and r = 0.1. So 0.9… = 0.9 / (1 - 0.1)
= 1
Intuitively, the prefix sums of 0.9… are 0.9
0.99
0.999
0.9999
...
and it’s "clear" that the limit of this sequence at infinity is 1. And because 0.9… is by definition that limit, 0.9… = 1.- Quoted phrase in the opening sentence
- Opening quote: “friend”
- Quoted phrase in the opening sentence
- Opening quote: “you are strongly encouraged to explore the use of AI”
College education for software developers has never been particularly good, but it seems like now they've actually just given up.
Do we accept this in any other field? Does your math professor teach you to solve integrals by asking WolframAlpha?
- Quoted phrase in the opening sentence
- Opening quote: “simple”
Actually, Feynman himself is a good example of that as well. His explanations are rarely short, but very simple. He has a series of YouTube videos where he explains all kinds of concepts like fire, but it still takes him at least 15 minutes to explain it.
So I’d argue that thinking it needs to be condensed into one or two sentences is getting the meaning of the expression wrong.
There are things where you may be given an explanation in simple terms, but the concept requires effort on the part of the recipient of the explanation to really understand. I’ll give an example: compactness. Probably anyone who has done analysis is familiar with how this concept is usually taught. Typically how it goes is you do open and closed sets and then they hit you with “a set is compact if and only if every open cover has a finite subcover”. Now this definition is extremely profound but very dense and it’s highly non-obvious why that in particular is the property you want (a finite subcover for every open cover)[1], how you could go about proving that every possible open cover in a certain situation could have a finite subcover etc etc.
In this case it’s actually quite easy to give simple explanations of the separate parts of the definition (open and closed sets, what an open cover therefore means, what a finite subcover then is etc) but actually understanding the concept of compactness requires something above and beyond whatever explanation you are given: it requires the reader to get out pencil and paper, try out some sets, go for some walks, think about it in the bath etc.
You may get a really fantastic explanation in simple terms from someone who deeply understands the topic, but if you don’t put in the effort you’re not going to understand it yourself. That’s not the fault of the explanation. It’s not an easy concept to grasp so you have to climb the ladder of understanding. The explainer can’t do that for you.
[1] it took the mathematical community years to settle on this being the definition, so you shouldn’t expect to understand why this is the definition straight away.
Feynman said "if you can't explain it on a freshmen level". Leaving out that qualifier distorts the statement.
The other quote about the magnets is also overly reductive. Saying "I don't understand it in terms you're more familiar with" is not an admission that scientists as a group don't understand it at all.
You could go into what it means to understand something, but unfortunately the article isn't doing that.
And then there's another barrier, of course: the audience. You may use simple terms, and tell a lengthy story, but some things are hard to understand. Relativity and Monty Hall come to mind. They can be explained in simple terms, but how many people get them? Probably just the ones that also understand a detailed explanation.
"Everyone who understands X can explain it in simple terms" =/= "Anything explained in simple terms is done so by someone who understands it"
- Quoted phrase in the opening sentence
- Opening quote: “works most of the time”
I think that describes the entire lifespan of consumer computing.
I’m old enough to remember Windows 98 running on a Gateway PC. That cost $2000 (without taking inflation into account!) and worked much less of the time than today’s OSes do. Modern OSes are frustrating in various ways but I think it’s easy to take for granted how stable they are compared to the past.
- Ordinary term used more than once in the same comment
- “good UX” appears 3 times
AI is not equipped for this. Humans are part of the experience process, and AI doesn't "understand" individual humans better than other humans do.
UX relies very much on human psychology. That's why, despite all the efforts of scoring it, UX is not an objective metric:
1. AI doesn't know what "good UX" is, as this is judged by the sum of humans using it. --> Structured data of actual usage would be needed for AI to ingest, in a way that it can derive intent from every action and identify friction.
2. A "good UX" is not a transferrable metric. The best UX for a ticket vending machine is not applicable to video-editing software. --> AI would need to understand what to apply when, and at which point it needs to evolve its expectation to the user
3. A "good UX" is not an absolute metric. Photoshop/Premiere/... has a UX which was extended and refined over decades, it is considered good by experienced users but still overwhelmingly confusing for beginners. --> Someone would need to scope the AI for the correct target-group
All this is done today by doing extensive user-testing with actual humans. There are people experienced in this who can create better UX-designs from scratch than others can, and still we keep learning something when we put a group of humans in front of it and watch them use it.
"Intelligence" alone doesn't solve this.
I'm reminded of the story of Stephen Hawking's sister, telling how he invented a "better board game" as a child and always wanted to play it with her, which was apparently the most frustrating and boring game she ever played...
As of now, I don’t see any of that being the goals even let alone the result. Instead what I see is that everyone is in a race to have as much product out as possible by the same standards that applied to products made by humans, but even lesser quality.
I the last two decades we have gotten so used to average or less than average product quality that we are willing to pay 2000$ for a device running an operating system like macOS or windows that ‘works most of the time’.
I hope we have products thanks to AI that compete on reliability of software and hardware to the point that the big tech will have to reinvent their processes just to be able to justify prices.
Thank you for saying this. It’s been on my mind a lot since the revelations of Huang’s affairs. When I saw a picture of his wife, who looks like a charming and nerdy woman who also looks her age, and then saw photos of the people he’s been sleeping with (younger, more attractive) it was rather heartbreaking. Building a company is incredibly absorbing. But when we are in our darkest moments - ill, or on our deathbeds, or after the loss of someone we love - it is not going to be the company who gives us comfort. It is going to be the wife or husband who has been with us through thick and thin.
Or not, because we betrayed that person, so they left. Because we forgot that “love and kindness are the foundations of a happy and contented life”.
I'm starting to get annoyed about the OMG OMG BREAKING BREAKING ALARM-EMOJI crap that's not a real emergency.
Four times in the last month, I dropped everything and started investigating whether a serious CVE related to any of our dependencies warranted an emergency bump, only to be frustrated that images, even from AWS, did not have the fix yet. So then I had to worry about whether I should build from source. We actually built PHP from source and deployed it to prod, just to get ahead of the PHP2Shell vulnerability
But yesterday's PHP vulnerability, CVE-2026-115560, didn't even apply to us or most people running PHP, and I don't think there are any realistic working exploits out there. I'm not saying that it's not possible, but I'm starting to feel that there's a large difference between what's theoretically possible and what's practically possible. And given the storm of patches and vulnerabilities found, this will become increasingly important to ask.
I think everything above a CVE score of 9.0 should warrant some level of “how exploitable is this in real life,” and anything above 9.5 should be proven to be easy for an attacker to execute, not just “possible in theory.” I don't see how we will survive as developers if there isn't some sort of better triage for vulnerability severity assessment. If everything is an emergency, then nothing is an emergency.
For those wondering about TLS: speaking as someone who has written a TLS implementation, AFAIK neither base64 encoding nor decoding occurs in the protocol itself, and the longest length field is 24 bits (Hello stream). Certificate processing involves ASN.1 which does support arbitrarily-long integers, but that's a big hint to check that they're within a sensical range.
You might find more people believing that AGI is already solved. For example, the author of "The AGI chronicles" drops this in the epilogue. For him, AGI is achieved, because it can help him with a wide range of subjects, with the ability of a random stranger. It's clear that he means chatbots, internet surveys summaries, and coding bots - he does not even mention self-driving cars there.
This is a shift from the earlier chapters, where AGI term is employed somewhat more vaguely, and even quipping once that AI researchers think of AGI as a slightly smarter version of themselves.
I too, have a lower bar for AGI. It would be enough for me if it could design games I work on. Merely speeding up implementation tasks is not that interesting. In the immortal words, a delayed game is eventually good, but a rushed game is forever bad.
I keep running into people who “just commit stuff” and they don’t know what they are doing.
Downside is we are heading into direction where people don’t even read code anymore so I might be just old man screaming at the clouds or agents.
I bought a new phone, GrapheneOS is on it now. I started moving things there one by one.
I got a home server and started self hosting my images and my docs.
You don't need to "do it all" just do some, it feels great.
fluidcad appears to be a remarkably good option now - particularly for something that only started development about a year ago.
I have used freecad, solvespace and onshape for at least a hundred hours each to make some reasonably complex parts. fluidcad resolves pretty much everything I found painful about those options.
I have some more limited experience with build123d which is also nice, but just not as easy and interactive as fluidcad.
Fluidcad provides normal interactive 3d cad interface with tools like extrudes/revolves/sketches but explicitly generates code as an intermediate step to constructing geometry and the code can be edited and refactored easily.
The linking between the code and the gui works remarkably well in practice.
It’s far easier to manage parameters and constraints in fluidcad where they are written out explicitly and can be edited vs what happened in a gui-only interface like Onshape.
If you’re playing around with a semi-constrained sketch to understand kinematics onshape tends to “helpfully” add random constraints which is really annoying to identify and correct. In fluidcad everything is way more explicit and visible and clear.
I have only spent about 5 hours using fluidcad but it looks remarkably useful despite some minor rough edges. It’s pretty much exactly what I wanted onshape or build123d to be.
Solvespace is a really beautiful piece of software but quite limited. fluidcad has the same sort of nice vibes as solvespace but is way more powerful.
- Quoted phrase in the opening sentence
- Opening quote: “test”
- Opening sentence uses a quote to challenge or distance a term
- Opening sentence challenges the term “test”
At a bare minimum, you will need to have a mass large enough to be seen on CT scan, and then an endoscopic biopsy (among GI scopes, the endoscopic retrograde cholangiopancreatography or ERCP is one of the most difficult). And even that may not get enough tissue for a diagnosis.
So, yeah, failure to diagnose it is common - largely because it's almost impossible to diagnose early. I'm a doctor. We feel sick when these things happen. It's horrible for our patients.
Jobs had a PanNET tumor[0]. That's the pancreatic cancer you "want". The survival rates are incomparable (iirc localized PanNET survival rate is around 95%!). Jobs chose woowoo bullshit over science, until it was too late. With his money he would have almost certainly been in the surviving group.
Edit to add: Generally when people refer to "pancreatic cancer", they are almost exclusively referring to Adenocarcinoma - the deadly one. That's what Swayze had.
[0] https://en.wikipedia.org/wiki/Pancreatic_neuroendocrine_tumo...
The speed is the biggest win IMO. When Carbon came out, I tried to give it a shot, but it was so slow that I went back to Opus just for productivity, nevermind quality. Today I started Neon on one worktree and Opus on another, and Neon did so well that I have now switched for both.
Code quality looks great. Comment-wise it has its own Claudese(exclamations like "this function has a concurrency bug and it's a sharp one!"), but overall I like reading it more than Opus. Might just be because I am not yet tired of its quirks though.
Oh, after a productive day on this particular work I had only spent $30 of Neon vs. $70 of Opus yesterday. YMMV
The fact that it's a monitor is a little bit irrelevant.