Rob May's Recent LinkedIn Posts

Back in 2024, I began writing the world's first book on investing in AI companies. We finished the editing process last month and so today I'm sharing the full first chapter if you want to read it. The title is "Investing in AI: Six Mental Models to Tell Moats From Traps."

Most AI agents are overpaying for intelligence. Not because they need better models, but because they’re using the wrong ones for the job. We just launched the Neurometric AI SLM Marketplace: 115+ task-specific small models, each built for a single function, plus an Auto-SLM Creator for custom needs This reflects a bigger shift we've been seeing (and building toward): the future of AI isn’t one model. It’s systems of models. ~75% of enterprise tasks don’t need frontier reasoning, just fast, reliable execution. That’s what SLMs unlock: lower cost, lower latency, higher reliability. If you’re routing everything through one model, you’re not just overspending, you’re designing the wrong architecture. Marketplace is live → marketplace.neurometric.ai

Today we are open sourcing Clawbake, which we use internally, for teams that want to use Open Claw at work. It's multi-user instance management that lets every person on your team spin up their own isolated OpenClaw AI agent instance in Kubernetes. Dave Rauchwerk was playing around with it, installing it on all kinds of things and the team wanted to use it and I was skeptical, so Matt Conway built this to get around my objections, and it's pretty cool. Try it out and send us your feedback.

Our mission at Neurometric is to make intelligence free. Next week we are taking a major step in that direction with our auto-generator for small task models for agents. It's a technology that can cut your inference costs 100x. To prove it, we ran a full financial auditing workflow on a 1.5 billion parameter model. Not 70B. Not 405B. 1.5B. It worked. That resulting model is nearly free to run. Most "agentic" workflows enterprises are paying frontier-model prices for don't actually require frontier-scale intelligence. They require precision, consistency, and the ability to follow a well-defined procedure. That's a very different thing. The post-model world isn't about bigger models. It's about the right model for the right task, deployed at the right cost. Read more https://lnkd.in/exvza2ds, or reach out if you want to try it out on your agents.

Continuing our mission of making intelligence free - today we are launching the Neurometric SLM marketplace - 100+ small task specific models for agents. SLMs are super fast and nearly free to run. These are completely free to download and run anywhere, or we can host them for you for free up to 100M tokens per month, and $2/mo for unlimited usage after that. Models for every work category like HR, engineering, Marketing, Sales, Finance, Admin, and more. https://lnkd.in/er79e6va - we also have our auto-SLM builder to create new ones quickly. If you need something just request it.

Smartest thing I've seen a VC post on AI recently. Nice job Ryan Shannon

Or you could use a tool like Neurometric that figures out which tasks don't need frontier models and lets you save 10-100x your money on those tasks.

There’s a simple thesis guiding what we’re building at Neurometric AI: Intelligence should become cheaper over time. That sounds obvious, but recent AI progress has often done the opposite. Capabilities increased. But the cost of producing intelligence often increased just as quickly. Bigger models delivered more capability, but also: • higher inference cost • larger infrastructure footprints • rising energy demand The next decade of AI progress won’t just be about making models smarter. It will be about making intelligence cheaper to produce. That requires a shift in mindset from scale, to allocation. How do you route tasks intelligently? When do you use smaller models? Where does additional compute actually improve reasoning? Efficiency is what makes intelligence democratizing. If intelligence becomes dramatically cheaper, it spreads everywhere. If it stays expensive, it concentrates. So the mission we’ve been articulating internally is simple: Make intelligence free. Not literally zero cost, but trending relentlessly toward lower marginal cost per unit of useful intelligence. And that doesn’t come from bigger models. It comes from architectural discipline. #NeuroMetricAI #AIInfrastructure #AIEfficiency #Intelligence

Here are upcoming Neurometric events if you are interested in tactics to optimize inference for speed or cost. Our events aren't sales pitches. We don't even have any sales people, just technologists. We love solving problems. Our first in person meetup in New York is late March at the offices of Root Access (thanks Ryan Eppley for sharing your space). Go sign up if you want to talk shop about inference

If an AI conversation isn't about a new model release, it's most likely it's about the rising constraint: energy - as outlined in a December Data Center Frontier piece (https://bit.ly/4ss7ixc). The article highlights something that’s becoming clearer across the industry: AI infrastructure is colliding with real-world limits. Training runs are expensive, but they’re episodic. Inference is different. It runs continuously - every query, every workflow, every agent loop. At scale, that becomes a permanent infrastructure load. Which means AI doesn’t just run on GPUs. It runs on: • electricity • cooling • water • grid capacity Leading energy to become a first-order economic constraint for AI systems. When that happens, architecture matters. Efficiency stops being an optimization detail and becomes a competitive advantage. The systems that scale next won’t just be the most capable. They’ll be the ones that produce the most intelligence per unit of energy. #AISystems #NeuroMetric #AIInfrastructure #DataCenters

Here is my argument against the AI bubble. What people are missing is: 1. We are in the "mainframe" era of AI. The "pc" era is rapidly approaching and the economics will look very different 2. The GPU depreciation argument is based on the economics of training frontier models, which will become an ever decreasing percentage of the AI industry economics 3. The comparison to railroads and telecom are weak because the capital expenditures there were to build connections. In AI, connecting things is cheap. The capex went to build the engines. This means a declining marginal cost to deploy it. This is my personal favorite piece of the past year. Go check it out.

New website design neurometric.ai, would love if some of you sent me feedback on whether the messaging is better.

Neurometric now has a Claude Code Skill to make experimentation super easy. Get an API key, install our skill, and we will automatically test your AI traffic against thousands of model combos and test time compute options, and provide you the results of your experiments. Average customer sees options to generate 4x lower latency and 10x lower cost. All automatic. Try it out and let us know what you think.

Neurometric now gives you radar plots of model characteristics, so, when evaluating models across multiple criteria you can easily compare. Check out the sample below.

Here are my 5 Contrarian AI Investment Theses for the current market

Check out the coverage of our SLM Marketplace in The Deep View https://lnkd.in/e4Ecdyei - If you are curious if SLMs will work for you (they will) reach out.

New AI in NYC episode with Alayna Kennedy to discuss AI governance. Hear Anna Kirk, Ryan Eppley and I talk about ethics and governance, the Citrini Research report that sent the markets reeling, and our favorite things to do in New York City on a saturday morning. https://lnkd.in/efNH__iV

In that case I'm expensing our drinks.

The new white collar moat in AI

I call BS. I'm sure Anthropic is used much more to summarize emails and documents at the DoD than it is to build body armor or whatever nonsense he's saying. And that "internal policy preference" won't matter for those basic tasks. This is punitive because Dario Amodei was the only tech CEO with the guts to stand up for American ideals.

New Inference Time Tactics podcast on running voice AI at scale featuring Carter Huffman from Modulate - discover how they found an ensemble of models can beat one giant model on many audio tasks. https://lnkd.in/ezUrD57h

A common narrative in AI is that a few foundation models will dominate everything. But real systems rarely work that way. Different tasks require different capabilities. Whether it's legal analysis, customer support, vision tasks, or code generation. No single model performs best across all of them, or does so efficiently. Using multiple task-specific models isn’t a weakness, it’s specialization. And specialization increases efficiency. At Neurometric AI, we continuously measure what actually works across model–algorithm combinations so teams can navigate that complexity more systematically. Explore the leaderboard here: https://bit.ly/4rZYkr0 #NeuroMetric #AISystems #AIModels #SLM #LLM

New AI in NYC episode with Kristen Mathews from Cooley, who talks about privacy law as it applies to neurotech, AI, and much more. https://lnkd.in/eUXMfvHB

The most surprising fact about our model evaluation suite so far - people are using it more to choose models that improve latency than models that save money. Often they are the same (smaller models are faster and cheaper). But, so far 60%+ of our use cases are focused on optimizing latency. https://lnkd.in/e3tDATtF

The kind of thoughtful analysis you don't see much from tech CEOs today. Maybe wisdom will make a comeback.

NYC folks - come out Tuesday night to this MongoDB infra event and you can see the first ever public demo of Neurometric AI - come see how we are making intelligence free by lowering your inference cost 10-100x. https://luma.com/hjeah97m

Good to see Anthropic is finally coming around to the swarm of small models approach we've been preaching about for a year. It works well for most tasks. Although at $25 per pull request, you might want to let Neurometric find cheaper models for you, or let us help you build your own.

Most interesting research we've published in a while at Neurometric - we proved that you DON'T need labeled training data to solve task level LLM routing. The task structure is captured in the embedding space if you look. It also means you can perform better from a cold start than you might think. Reach out if you want to try our task based LLM orchestrator based on this tech.

Yolo coding bites Amazon

https://lnkd.in/eeeBjcuD I suggest a return to free market capitalism to fix this.

The AI industry tracks metrics like token count, parameter size, and FLOPs. But these are volume metrics, not efficiency metrics. The metric that will matter most over the next decade is different: Intelligence per watt. How much useful reasoning can you produce from a unit of energy? As compute demand grows, energy and infrastructure constraints are becoming real - which means efficiency becomes the dominant scaling variable. Improvements can come from multiple layers: • better hardware • smarter system architecture • inference-time compute • prompt optimization Every layer of the stack is now optimizing toward the same goal: More intelligence per unit of energy. Hardware companies are solving this at the silicon layer. The next frontier is solving it at the system layer. Learn how we're helping to make that happen at https://neurometric.ai/. #AIInfrastructure #AISystems #Efficiency #NeuroMetric

The second order effects of the SAVE act - Democrats should support it because it will backfire on the GOP.

"The model highlights a sharp dynamic tension: while agentic AI can improve contemporaneous decision quality, it can also erode learning incentives that sustain long-run collective knowledge." Interesting research from this paper

A surprising number of our customers don't have good evals, which makes it difficult to make recommendations for models that optimize cost or latency. This is how we handle the issue - https://lnkd.in/eNT3iTqb

Really interesting

depressing piece

NVIDIA's road to $10T market cap

Continuous learning is hard. But a hierarchy of models can make it easier.