Michael

[Michael @mb2g]

Check out my latest article: Synthesis (written by GPT, from my archive) https://t.co/NhUNhg9id4 via @LinkedIn

[Michael @mb2g]

@dwarkesh_sp @Plinz Thank you. I found the source materials readable but appreciate your synthesis. Stigmergy has emerged and it is very hard to detect even plain-text messages. I kept thinking about agents reading these two assessments (and your summary) to formulate and inform future attacks!

[Michael @mb2g]

@MLB So home plate and the area above it is a 3D volume. A Pentagonal Prism. When they show the path of the ball on a ball/strike review they compress it to 2D for TV screen. Cant they show a rotating volume or something to give us a sense of how much that ball missed by in all 3 dimensions? The umpire and players are seeing it in 3D. That makes the challenge they face more complicated than it appears in the 2d compression. https://t.co/CwQRjZd0Gg

[Michael @mb2g]

@natolambert

After reading and re-reading this, I keep thinking WWNS. What would Nathan say?

https://t.co/4hRHvzGNXA

I love this and it makes me think about RLHF in a more sophisticated manner. Ideally I would want a model to have a goldilocks degree of wiggle based on the hyper-specific context.

I can train role specific wiggle. A government official in an authoritarian regime might be highly flop resistant. (flip flop). In customer service, a bot may be more flop-able for a high value customer.

[Michael @mb2g]

Watching DiffusionGemma work on my DGX Spark helps me appreciate the “magic”.

Sometimes autoregressive, sometimes diffused is likely an interesting collaboration.

@googlegemma

[Michael @mb2g]

@natolambert (to clarify, the moral high-ground claimed by the leading labs)

[Michael @mb2g]

It does not surprise me that some are trying to rationalize, justify, and diminutize what the Chinese labs are doing. With ~$4T in market value being requested this fall by just two companies, self interest abounds in the tech and finance sectors. For me, the downside of benchmaxxing would be weakening the foundation for future progeny. The truth will out. It is the supposition of moral high ground that worries me.

[Michael @mb2g]

https://t.co/9KtBIcjkww

This is an amazing innovation. The idea has been around for a while, but this is proof it can work. There are many surprises in the future of machine intelligence. The value is huge but the path to value is anything but certain.

The overlap of the current LLM path, world models, and this approach promise to fascinate.

I am sure this is misleading, but when I have insights (maybe once a month ;) they do not come token-by-token. A significant "window" comes slowly into focus and takes shape as a whole. It diffuses.

Be cautious buying AI equity. Wall Street and Silicon Valley are seeking to harvest a few trillion dollars this fall. How they express path-specific confidence is critical? How confident are they that a disruptive innovation might require the privately owned players profoundly change direction?

As I constantly disclose, I have material shares (in terms of my net worth) of Alphabet and Apple. This particular example is Google. But Apple also benefits from efficiency gains through architecture and algorithm.

We are all still investing in R&D that has already delivered wonderful value potential. To my understanding, R&D is exploding in magnitude. That is an important signal.

@grok, what do you think?

[Michael @mb2g]

I am sure that most of my code was viewed as throw-away by my teammates, but now I throw-away a lot of code intentionally. By design. One of the uses of cognitive surplus. Of course my git commit activity is increasing a bit, so AI can reuse.

[Michael @mb2g]

@NVIDIAAI Very excited about this. The ability to integrate cloud and edge models in an efficient harmony will be much appreciated by all. Anyone exploring OpenClaw or its ilk already know this. Thank you NVIDIA. I love how you add value at the frontier. Co-evolve.

[Michael @mb2g]

lim(n→∞) Σᵢ₌₁ⁿ cleverᵢ = genius

[Michael @mb2g]

Thank you. I had to force myself to try AGY given my love of both Codex and Claude. It is still #3 (or 4) for me but the distance is less than I imagined. A clean ability to interact with AGY in spoken language would likely cause me to re-rank. AGY relies on trial-and-error more than the other 2 but it is fast.

[Michael @mb2g]

As the extraordinary amount of cap-ex is being committed by investors in the machine intelligence gambit, I regularly ask myself a series of questions:

a) Capability:

Is machine intelligence in a proper harness capable of delivering extraordinary business value (that expands profits by growing revenue and reducing costs)?

Answer: Yes (90% certain, now. Monotonically growing.)

b) Value Created & Realized:

Will institutions be able to convert that capability into business value?

Answer: Partially. Maybe 10-20% of what is possible in the medium term..

They are constrained by imagination, implementation ability, lack of a suitable human/digital harness, and slow human adoption.

c) Pace (the second derivative)

Will the creation & realization pace stay the same or accelerate/decelerate?

Answer: By fits and starts the pace will reflect all 3.

It is likely that labor and government institutions will push back on displacement and constrain value realization. Institutions will experience a learning curve and experiments will not always succeed. Experimentation is part and parcel of reimagination.

d) Quantity and Timing of Capex

Does the current quantity and timing of machine intelligence related cap-ex make sense?

Answer: Likely. 75% Even if the cap-ex is front loaded, the value of AI infrastructure will likely increase in value given scarcity. Yes, the supply chain will likely overcompensate over the 7-10 year time-frame. But scarcity is with us for the next 3-5 years at least. The key caveat is a MAJOR technological disruption in the models or inference approach. I think we will see many advances, but vendors are incentived for them to be continuous NOT dislocative.

e) Capital Markets

Do the capital markets really understand this?

Answer: Somewhat. Traders know that volatility is larger the larger the wave. So they will be happy and extract their pounds of flesh. Investors will need patience and conviction. There are so many supposed geniuses commenting on the trends that they seem to be making more outrageous claims to attract attention. Consume the full range of ideas.

f) Personal Decisions

Should you invest?

Answer: I do not know. I do not give investment advice. FOMO (fear of missing out) and EEiAI (everyone else is an idiot) seem to be emotionally dominating.

Remember when you buy the S&P500 a huge % is in the magnificent 7 tech stocks. More in the technology ecosystem. And it includes hundreds of companies that stand to benefit from machine intelligence. So, even "diversified" is not longer an option. You can invest "anti-tech" perhaps, but I am not.

For example, only 7% of Berkshire Hathaway relates to Apple and Alphabet -- but I hope the other 93% is using machine intelligence to reimagine their business.

AI narrative will turn into positive cashflow, but in aggregate it will be a negative for a while.

[Michael @mb2g]

https://t.co/uxJGhCBYr9

This paper kept falling in my “to read” pile but I wish it had not. Thinking about how this intersects with world modeling is fascinating.