cashsterling

I think David's expressed concerns are valid... there should be more safety engineering around LLM to avoid data loss and other accidental issues caused by LLM models.

And, I find the timing of "all of this" very interesting:

- we have a coordinated marketing campaign around discussions of AI safety. I am not saying David is part of that campaign but it would not surprise me. OpenAI and Anthropic are talking about safety and the need to slow down. People are "resigning". Techno-influencers are also being offered money to talk about AI safety, etc.... and some of them are taking that money and others are not.

- Cheaper models are keeping pace with Frontier models, albeit with a 3-6 month lag... which indicates that frontier labs have no moat unless one is created for them by law or some other market force.

- we have a ballooning debt crisis in LLM model hyperscale spending... revenue from serving LLM models is lagging the debt payments which are coming due by a wide margin. If something dramatic doesn't occur in the near- to mid-term, the financial bottom is going to fall out of this thing and the AI industry is going to suffer a major collapse of some sort.

LeCun is a lot more than a "Grey Beard"... lol, that is a significant understatement of his contributions to the field and position in the field.

Among other things, LeCun is one of the senior people industry who has a deep understanding of the mathematics and analysis underlying neural networks and "neural network like" approaches to machine learning. I think he understands a lot more about neural networks than most researchers today.

I also don't think our current LLM models are AGI and think the LLM approach is, mathematically, incapable of producing an AGI. At the end of the day, the architecture is still a streaming token plinko machine with a lot of guide-rails to achieve good behavior.

I do think it is very impressive how far hundreds of billions of dollars have been able to take LLMs in terms of usefulness (I use LLMs everyday in my job). AGI or not, current LLM models are a pretty incredible achievement and will provide lasting benefit even after the economic implosion of the AI industry occurs (which I think is imminent).

A lot of shopify's engineering probably work on the back-end business logic for payment processing, sales tax, etc. Shopify has to comply with financial norms, rule and laws in so many different jurisdictions.

I know someone who works for Gusto (HR and payroll SaaS for small businesses)... most of their code-base is also business logic and whatnot for all the jurisdictions they support.

I too doubted, from the beginning, that neural networks will be the basis of AGI. As impressive and useful as LLM's are, they are still a long, long way from AGI.

It is not a one dimensional take... it is a stress test of qubit gate fidelity [across all qubits involved in the circuit], state prep and measurement , lifetime (coherence), memory errors, etc.

Now I agree that there are other great stress tests of quantum computer systems... but most of the industry agreed that quantum volume was a great metric several years ago. As many companies systems have been unable to hit decent QV, companies have pivoted away from QV to other metrics... many of them are half baloney.

The better analogy here, in my opinion, is until someone builds a car that actually drives forward...

A lot of these quantum computing papers are just loud engine rev'ing.

But you are right... "zero attention" is hyperbole... I shouldn't have said "I pay zero attention"... I should have said, "I am underwhelmed and don't give a lot of credence to systems until they demonstrate ability to perform high fidelity calculations of reasonable complexity and depth."

I do think superconducting qubits approaches will continue to improve in fidelity and ability... there are just too many brilliant people working on these challenges to count them out.

Until a company can demonstrate a quantum volume of even just 2^16... their computer is just about worthless for any sort of half-real quantum computing.

I pay zero attention to any technical information and marketing speak coming from a quantum computing company until they demonstrate decent quantum volume.

Most companies computers can't even hit 2^16 so they are figuring out means of distracting the market from the poor fidelity of their systems.

Modern high-end fabs have extremely expensive equipment and are highly automated... like they are so automated that people don't actually handle wafers... it is almost all robotic.

Thus, salaries and cost of services do not factor in as heavily as you might think to fab economics.

Data suggests that TSMC's per wafer costs in Arizona are 10-30% higher than Taiwan and that Arizona fab is relatively new. It's economics will probably improve over time, narrowing the margin to 5-15%.

Looking towards the future, power costs and other global supply chain factors could very easily make TSMC's Arizona fab less expensive and more reliable to operate over time. For one, the US is completely energy independent... Taiwan is not.

Atlantic Quantum has some brilliant people and I'm sure they have some great ideas.

That said, I dislike the quantum computing craziness we're in where big claims are made without proof or data.

"We have the fastest clock speeds, lowest error rates, and most scalable architecture... no data, just take our word for it."

It will be interesting to see how they do in the DARPA QBI program.

I always try to communicate a strong sense of hope to my teenage kids. Many nations have an inverted population demographic, including the US (especially without immigration).

I think the US is "as little as 10 years" away from a significant skilled labor shortage.

I think Ada is a great language and it is completely possible that Ada will experience a resurgence in coming years.. especially as LLM's are used more and more to generate software. ADA/Spark can provide very robust guide rails for correctness (so can Rust) of 'AI' generated code.

Yeah... my intro CS class was in C and Ada95 (I'm not a CS guy btw, just took the class). I actually preferred Ada over C... but continued to program in C for other classes because of compiler availability; I had to do all my Ada programming on Sparc workstations at school.

I personally think that AdaCore, and friends, missed an opportunity in the early 2000's to fully embrace open source... they left a big gap which Rust has filled nicely.

I still think Ada is a great programming language. When a question comes up along the lines of: "what's the best programming language nobody's heard of?", or "what's the best programming language that is under used?" Ada is usually my first answer. I think other good answers include F#, <insert LISP flavor>, Odin, Prolog, Haxe, and Futhark (if you're into that sort of thing).

I really hope the best for Oxide and applaud their compensation model.

I applied to one of their roles which required me to write about 10 pages of text to answer all their questions... which I think is a big ask but I did it because "why not".

They took over 3 months to get back to me, but at least they got back to me (with an apology and a polite "no").

100% agree with you!

I have worked US manufacturing and manufacturing R&D for most of my career: pharmaceutical, microelectronics, materials, aerospace, etc. The US is awesome at manufacturing when we want to be.

One problem is that "modern MBA/business philosophy" views manufacturing and manufacturing employees as a cost center and there is so much emphasis on maximizing gross margin to increase shareholder value.

So business leaders scrutinize the hell out of anything that increases the cost of their cost centers:

- employee training & development? hell with that.

- Increasing pay to retain good employees in manufacturing? Why? isn't everything mostly automated?

- manufacturing technology development? Not unless you can show a clear and massive net present value on the investment... and, then, the answer is still no for no good reason. I have pitched internal manufacturing development investments where we conservatively estimated ~50% internal rate of return and the projects still didn't get funded.

There is also a belief that outsourcing is easy and business people are often horrible at predicting and assessing the total cost of outsourcing. I have been on teams doing "insource vs. outsource" trade studies and the amount of costs and risks that MBA decision makers don't think about in these situations really surprised me initially... but now I'm use to it.

Anyhow... the US (and Europe for that matter) can absolutely increase manufacturing. It is not "difficult"... but it would be a slow process. I think it is important to differentiate between difficulty and speed.

I graduated in ChemE in southern California in 1999 when there was a major downturn in the job market. One or two big chemical engineering design firms closed their SoCal offices flooding the market with qualified chemE's, aerospace companies were consolidating, etc.

Some of my school colleagues got good jobs at refineries and whatnot... but they were the fortunate ones. It took me 12 months to land my first "I made it" engineering job with a good salary. In the interim, I worked hourly jobs making between 13-18 USD an hour.

Don't let the current job market deflate you. You are young, intelligent, and you have a degree from MIT... you are going to be fine.