lumost

This, I happily used the opus 4.6 fast mode to the tune of 5k for a project. The delivery of the project justified the 5k, if I only spent 500 but delivered the project 1 month later - I would have been in the dog house.

In my opinion, this comes from the 70s and 80s where there was very real concern that Japan was going to surpass the US economically. Many companies in the US attempted to adopt Japanese methods in manufacturing and other areas, media then inherited further Japanisms. There is also an historic backdrop of westerners viewing Japan as a mysterious civilization on the far side of the globe dating back to the 1500s.

At the size of IPO courting investors is inevitable. You need to make sure that there is enough money available.

These companies are burning enough cash that they will need to be public.

We’re about to have 3 of the worlds’s largest corporations be massively in the red.

Napkin math on 5 year depreciation is 5.5 billion per year for 28 billion. However the 28 billion is cash upfront, spacex is probably paying 10-20% interest on the 28 billion for another 2-6 billion per year.

So net you are looking at finance expenses of 7-11 billion per year. The electricity costs will be significant on top of that, but harder to get a solid read on.

Net of everything, spacex may be getting a 14-28 percent yield before paying for electricity. After electricity/insurance/data/taxes/other expenses - I’d guess it’s anywhere between 0% and 7% yield.

Odds are good that Anthropic abandons the deal before the depreciation schedule completes. Who is going to rent the GPUs then?

Not my area either! But my understanding is that there are more efficient methods of representing static numbers when you can skip the vram lookup.

https://taalas.com/

Is an example startup in this area claiming 16k tok/s on an asic for llama 8b. Qwen has a 27b model at opus 4.5 quality.

Capex, opex, quality, and volume are tricky things to balance. On balance, pc/mobile are cheaper to operate than equivalent cloud and on prem deployments.

It’s not unreasonable to suppose that in 2 years time an opus 5 quality model will be etched into silicon for high performance local inference. Then you just upgrade your model every 2-3 years by upgrading your hardware.

This probably ends in a deflationary spiral. The ai replaces the jobs, the lack of jobs chills demand, the ai becomes cheaper because it exists in a commodity market.

The money printer will be used, and maybe it all works out - or we see wealth hyperinflation and build out our own aristocracy.

I find it interesting that Mythos was announced the same day that GLM overtook opus4.6 in capability. To me this seems like a careful attempt to cool demand for opensource models which are about to take the overall lead.

It's more complicated than this, The US has multiple challenge in its own domestic talent pipeline. In a world of finite slots for elite production and elite employment the US must own the outcome of allocating those slots internationally and the resulting under-employment of its domestic population.

Whether these slots should be finite or not is an independent problem, however for various reasons the slots are currently finite and potentially reducing in volume with income inequality.

The 9-5 is doing a major part of that comment. Irregular sleep isn’t the end of the world if you can sleep in and recover. Modern parents don’t get a chance to recover.

Honestly, the problem is one of BS detection.

Lead engineer says something is not workable? Pm overrides saying that Claude code could do it. Problems found months later at launch and now the engineers are on the hook.

New junior onboardee declares that their new vision is the best and gets management onto it cuz it’s trendy -> broken app.

It’s made collaboration nearly unbearable as you are beholden to the person with the lowest standards.

If the project fits on a single server (which it probably can in 2026) docker compose is great.

Are we really any different? :) you could easily extend the above story to consolidated images/video/3d environments etc.

Currently, it is difficult to live update the model’s parameters in response to new information. This difficulty applies at both an infrastructural level and an optimization level.

We simply don’t know how to incorporate new information without losing old capabilities reliably. Pans handle this through extensive evaluation, heuristics, and experience.

What we do know is that models can adapt to their context, and extending the context window is an infrastructure and capex problem first. A billion useful tokens would obviate the need for any out of band memory structures.

This gets a bit tricky. Over very long task contexts (1M tokens) or with prompt compression (10s of millions of tokens) the model can alter its priors based on updated evidence. This form of knowledge based learning is not necessarily robust, but demonstrably does occur.

the latent space of the LLM when it chooses each token is 10s or even hundreds of GB for each word that it chooses. It's not really useful to look at LLMs from the perspective of its prediction head which is a very small part of the model.

Having just been on the job market, my experience was that career pivots are much harder now. I initially intended to transition to a neighboring field after an education break - but none of the companies in that field would speak to me. At most I had a recruiter call with one who decided to reject. To complete this transition I had been planning on a massive pay cut.

When I focused on areas I had some more credible experience in, I got significantly better engagement and eventually found a very narrow niche where I had substantial success.

I think we're partially adjusting to a world where employers expect a very narrow experience match to their role. Employers are also paying a premium within that narrow match.

I spend 400-500 dollars per day during active development at this point. However with more aggressive task breakdowns I can spend ~5k per day.

These spend rates are in part due to operating on a larger code base. Operating on a larger code base means more time searching and understanding the code, tests, test output. They are also due to going all-in on agentic coding.

It can feel painfully slow to go back to coding by hand when for a dollar you can build the same functionality in a minute. Now do this with multiple sessions and you can see where the cost goes.

I'm sorry, but what kind of CalvinBall allows this argument to be made in good faith?

https://www.whitehouse.gov/presidential-actions/executive-or...

I had Claude dig through the orders to pick out substantive examples where the courts ultimately allowed the order to go through USAID's dismantlement is the top instance which has an affirmative legal ruling. However unilateral grant freezes, tariffs, and other issues are still progressing in the courts.

While I'd certainly like to return to the world where congress handled the democratic duties of law making, budget, and war declarations. I must acknowledge that we no longer appear to have separation of powers.

This sounds like an issue where the hyperscalers are acknowledging that the new Foundation model firms may in fact be worth more than they are. Anthropic looks increasingly likely to exceed AWS revenue next year, and OpenAI will likely do the same with Azure.

3 years ago a Foundation model seemed like a feature of a hyper scaler, now hyper scalers look like part of the supply chain.

Are they really? perhaps it is just my social circle - but I've observed most of the group traffic having moved to SMS and WhatsApp. While WhatsApp is a meta product, the economics are extremely different.

There are really just a handful of people from my extended social circles still actively posting to FB/Instagram and that traffic is mostly drowned out by slop content.

Meta has no real product moats at this point. Yes, many of us still use instagram and facebook on occassion - but I'm not giving it any data beyond "what short form slop content will I accept".

The days of meta having network effects to defend its position are long gone, and I suspect we'll see the products die when an AI-first UX comes to mobile.

Out of curiousity why do you not refill tokens in this case? When I'm actively working on a project I'm prone to spending a few hundred dollars per day or a few thousand during the initial buildout of a new module etc.

The same pattern will also play out with Universities and colleges. 2025 was the year of peak US high school graduation, with the next ~30-50 years of graduation rate declines baked in. We're still a few years away from this trend percolating into the work force.

Boston Public Schools had a 50 Million dollar budget shortfall for next year. We are rapidly closing schools and eating the disruption that comes with that. Teachers do not do their best work when they don't have confidence on long-term outcomes.

To some extent, this shift is inevitable due to demographics changes - but I don't think that there has been realistic planning on how to manage a future with dramatically fewer children.

To add to this, Cursor provides a high value go to market strategy for X.AI's modeling efforts. Cursor's own modeling efforts would require an extreme investment of capital to compete. Capital which X.AI has already spent or is planning to spend.

A lot of organizations live in some game theoretic equilibrium that prevents cost improvements from being metabolized by the org without burning the cost elsewhere.

For example, consider a commodity business for software product X. All vendors of this product had their costs reduced by a factor of 100 over night for developing new product. They could increase their profits, lower their price, or re-invest the dividend. In software, the buyer usually buys on quality - so they all re-invest.

Now they are spending the same amount on product development, for the same price tag, and earning the same profit - but they might be shipping much faster.

industrialization gave us cheaper goods, and new types of goods (the automobile). What new consumer features has ai delivered? So far, I've seen mostly slop.

Codex is my favored coding agent for generic "I need an agent tasks." GPT-5.4 does a bit better with images compared to claude, and debugs a little bit better.

The UX of codex is exceptionally nice however.