lumost

I think the interesting question is whether the average consumer cares? a 4B model should hit roughly the same numbers as GPT5.5 sometime in December/March. That is more than capable of performing a variety of complex work tasks.

The margins on NVidia datacenter hardware are ... high. At least one order of magnitude larger than a consumer chip.

Given the recent deepseekv4.1 advances - how good of a 3B model can we make to run on an iphone natively? is it good enough to match common muse/dot use cases for consumers? the phone is already always on.. no need for a cloud server.

It would be interesting to see how these are justified internally. If I had to guess, it’s either excess budget being channeled into “something” or a justification to reduce street patrols/street crime.

Out of curiosity couldn’t this also mean that we are missing the agriculture? Or that the practices they used were very different from today.

We really need more slow charging stations for folks to just park at. I’d have assumed that the economics would have been favorable to charge by the day/night rather than by the minute.

AI has largely not been trained to operate independently yet. Newer products like grok bot/meta muse point in the direction of an asynchronous "virtual worker" type of persona where the AI is empowered to make at least some decisions.

It’s worth recalling that a large (possibly majority) share of the population has extreme limits on their time imposed by economics.

While I’m fortunate now to be able to dedicate time to the gym and other physical fitness, this was not always the case.

We lack the societal incentive structure to allow someone working 60-80 hours a week to reduce their workload to physically recover. We do allow that same person to spend 5-10 years income on surgeries which have 4-6 week recovery times.

How confident are we that they know it’s fake? I somehow doubt they close the door and pull out the real information.

I suspect that we could bring this measure into the modern world with a little help from either DFS or ai.

Something like abstractions traversed during interpretation, lines of abstraction v.s. functional implementation, or logic statement dispersion.

It was hard to pin down what was abstraction vs. implementation, but it's much easier now.

Biology is very messy and very data constrained. We think that human anatomy works one way- but surgery and radiology are hard because every human is very different.

Missing organs, odd tissue performing unclear functions, different, new or missing muscles.

It’s unclear whether a world ending virus is even possible. There might be better luck with prions, or fungi.

It's fairly obvious that the frontier labs are proposing that only "paced" providers would have access to the US market - and by extension the G7 market. They don't mind splitting market share between the three of them, they do mind open models destroying the "frontier lab" operating model.

Even if all GCR models were banned, anyone with access to 20k-1MM GPUs can throw their hat in the ring to make a new frontier open model.

While I dislike this take, when the dust settles we will have compute clusters orders of magnitude larger than anything that existed in 2023.

If demand vanishes for the 3 million cards Amazon just bought, then something will be done with them. The AI market may end up in a bizarre jepson's paradox of rotation between inference use cases and model training.

While simultaneously trying to own all economic activity derivable from labor... it's hard to see the argument as anything but disengenuous.

Maximally, a closed form solution would remove the need for Computational Fluid Dynamics. Any property could be derived from a (presumably expensive) analytic function.

Minimally. It could say that no such analytic function can ever exist. Which would be rather boring.

Turbulent fluids look awfully predictable with their spirals….

How do you know the lean is correct? You don’t bet the two trillion dollar company on “the ai said so”

The loss of a succession strategy with an elderly leader at the helm seems like a major issue.

I can do none of those things.. I hope that I am generally intelligent.

1 year ago we viewed models as tools and agents were just kinda toying around, that we now think the bar is literally an anything to anything converter through one agent is wild.

I think we're getting to the point where it is difficult to identify the goal post of AGI.

Is it rapid skill acquisition? -> ARC benchmarks are saturated Is it breadth of knowledge? -> See many ... many benchmarks Is it ability to do hard tasks? -> see terminal-bench and released outputs.

We are at the point where the starting point for most tasks should be "send your agent to work on it."

So where do we draw the line in a way that doesn't move every 6 months?

Then why does the expensive treatment cost 7 figures? Estimating hospital costs is hard, but a quick glance at many buildings/offices would tell you that they are by no means struggling.

We overpay for healthcare because we must pay, and it has institutional and natural bottlenecks which restrict availability and choice.

A lot of big enterprises struggle with ai adoption. Either they can't get models/tools approved fast enough, can't provision them in their internal network, or get slammed with exorbitant inference costs.

A big enterprise can drop one (or 4) of these on someone's desk and let them go nuts.

There are many enterprise environments where running modern models is... difficult. Rather than fighting security for months, a user could petition for a mac studio and have rough cost parity with a z.ai subscription.

if they are a heavy user, perhaps they string 4x together.

As a kick-starter project is this thing actually real? My experience with kick-starter has been that the projects are almost universally faked.

The answer is likely that it makes sense for maintenance purposes, but the legal authority granted to an HOA is overly broad.

In the past, a professional job might mean extended periods of travel by rail etc with little work to do. A small acreage farmer has to wait for plants to grow.

The advent of factories with limitless and constant demand for labor is a new development of the industrial revolution.

Token cost is falling rapidly for a given quality. We don't think about this too much as newer models have made legacy apps obsolete - but its an interesting question what the cheapest text to sql or similar model would be. I still go by 10x cost decrease per 6 months for any given model quality.

betting on a compute bottleneck sounds like a recipe to get thrashed when the bottleneck relieves itself.

At the investment scales being discussed, CUDA/architecture and other advantages do not matter - you could spend 1 billion on building a new chip architecture. The ram/fab inputs have been a commodity market for years. Heck, even the model bottleneck doesn't seem real when it's only 1-4 billion or less to get a state of the art model.

At some point the compute bottleneck will be relieved, you can see NVidia hedging their strategy with both open models and on-device chips targeted for local inference. The 200 dollar a month plan will absolutely be taken over by local hardware in the future.

why would anybody believe true brilliance? They would only believe something that can be proved from combining existing knowledge.

If a time traveler went back to 1600 and started spouting off about differential equations everyone would think them quite mad.

I'm still uncertain if mythos is real. Subsequent model releases have been lackluster, no one has claimed to verify mythos performance and it's silently vanished from most comparisons.

I would presume the goal is to have an electromagnetic accelerator of inbound charged ions to serve as thrust. You can get limitless power from solar, but propellent mass is a limiter.

It seems like a data/roadmap acquisition. Reports have been that Cursor's footprint was still growing rapidly, and their composer model wasn't bad.

Take that data / training recipe and scale it up with compute to see what happens?