@OfficialLoganK@GoogleAIStudio Google making everyone a chemist trying to figure out how the model names map to model sizes
H
: -> C -> Ne
:
Ar
Li next? Is Si the holy grail?
deployment is a multi trillion dollar blindspot for investors.
i spent a couple days on panel speaking about this. all of the investors attending are interested in hard tech.
capital markets are staring down the barrel of trillion-dollar frontier model valuations. the prevailing estimate is that we are currently at one percent of the compute we would have deployed by 2040. if you believe this, being a buyer of foundational ai is the bet you would make.
aside from that estimate of exponential compute, there is also quiet concern amongst investors looking at the space: "where is the actual enterprise adoption?"
across markets, pilots are stalling out. nobody is saying out loud that most early ai initiatives were not intended to create ROI, simply a broader shareholder signal campaign. but in hard tech and industrials, performative pilots are not exactly a priority amongst the massive backlogs and supply chain issues.
investment in hard tech is undergoing a pivot. investors are realizing that this is a distribution problem. venture returns depend upon scalability; this is where it gets hairy.
walk onto the factory floor of any major manufacturer and you discover this: everything runs on clipboards, operator knowledge, and not touching things that work. VCs are evaluating hard tech startups as if the customer is ready for physical ai. they just aren't. aside from that, manufacturers don't buy pilots running inside non-deterministic boxes. they have already been burned by the empty ROI of IoT and industry 4.0.
here is the blind spot. the true opportunity over the next 5 years is building the three missing components: data acquisition, physical actuation, and deployed engineering.
let's dive into the most misunderstood category. deployed engineering has been seen for years as unscalable. vcs look down on it as a form of consulting with poor scalable margins. this is going to change.
having previously run a company that trained 5,000 engineers, i knew this had to change. the role of the next decade is a forward-deployed engineer. why? most system integrators depend on effective allocation of engineering hours for their bottom line. ai should, could, and likely will leverage that allocation.
factories are like complex codebases. you cannot sell a module into the stack without engineers moving things around to accommodate it. the issue is even more serious in the physical world because the changes have permanence. it cannot be rolled back, cut over, or recovered.
for physical ai companies looking to deploy, you will learn this: every deployment you sell will be bespoke. there is not an immediate world where the adaptability of the product creates a plug-and-play setup experience. this is compounded by the other issues: capex, insurance, maintenance, operator harmony.
so, as investors calibrate for this reality, they will be filtering for a few things:
-tools that make engineer output higher
-processes that make engineering organizations more effective
-companies that have mastered the pilot-to-deployment gap
many, many valuations ignore this issue. tech potential is great, but for the 15-trillion-dollar manufacturing market, there must be real use cases that stand up.
Ted talks a bit about materials science discovery here. Even if we scale that to 100s to 1000s of experiments per day, scaling up a promising discovery to manufacturing scale puts us increasingly expensive and hard to acquire data regimes. This is where most deep tech discoveries die.
Today this scale up is mostly an art, driven by intuition, locked or forgotten inside large cos or in places like China who have been scaling deep tech for years. Totally inaccessible to most humans (and to ai).
Best talk Iโve seen in a long time and the clearest articulation of where the bottlenecks and the opportunities to build really are in the physical world. Couldnโt agree more!
Was so great to discuss this with @xiao_ted and @willsentance in person a few days ago
Iโve spent a lot of time thinking about what a post-AGI world looks like. That world is rapidly approaching, and it is useful to consider what the most pressing challenges and opportunities will be. I recently shared these ideas in a talk, which should be quite digestible.
@JasonrShuman 100% agree with this. First step is really deeply understanding customer workflows and what is actually worth automating. Your post reminds me of Elonโs principles on automation.
Many robots have run on two models: a slow one that decides and a fast one that moves. That split is dissolving. What survives is the layer that keeps a robot from falling over, and a check on what it plans to do. https://t.co/WcqdgyrQtb
Lots of parallels to OAI and Gemini live voice models and thinking machines interaction models. Kind of like how my dot can continue the convo at low latency and offload heavy duty tasks. Also the fast models will become stronger reasoners over time so the number of offloads will decrease like you say.
In a world where robotics gets increasingly generalized should industrial companies be building up โinstitutional contextโ that will help these robots succeed?
What should this look like and what lesssons from LLMs translate? Which do not?