If you dig a bit youāll find it says R&D CTO on my LinkedIn profile. What does this even mean?
Chief Technology Officer - this means I lead the technological aspects of startups, and this is a natural companion to the staff engineer role Iāve had with bigger companies.
R&D - most CTOs are operations focused, managing staff and budgets and such. When Iāve worked in organizations that need that, Iāve been in the staff engineer role. With startups itās all about my ability to digest the new and pick a course of action, usually well before most others understand what is happening.
Sunday I got to do something for my day job that involved Parabeagle, the legal/research fork I did of Chromaās MCP server. Since this is high level strategy stuff that does not expose what Iām doing, I can show you guys the details.
Attention Conservation Notice:
Everything is moving way too fast. You should depend on good sources in this environment. You can use AI, but LLMs both lag and lie, so itās better if you can use them to get answers from curated expert material. If youāre not running MCP servers on your own, this will be a bit much.
Motivation:
Nate B. Jones did a review of the State Of AI report for 2025. If you want to know whatās happening, here on the edge of the AI bubble collapsing, better watch it all.
The things that got my attention the most were the environmental and geopolitical aspects. Weāre at a point where the problems the sector faces are not due to bytes, theyāre due to atoms. Electricity and water are issues, supply chains are issues, and where things are made versus where they are used is a rising issue. And thatās before we factor in the as yet theoretical Artificial General Intelligence.
Methods:
I watched the video. The next time I watch it I will have reviewed the State Of AI slide deck he is summarizing and Iāll have a pad & pencil handy. Not electronics, a legit beat up tablet and a mechanical pencil.
But I have associates who need to know these things, and they have not enough time, or they will not take information via audio/video, and since I need to understand even more broadly than just the 2025 report, I started curating. Specifically:
Transcript of the video.
The 2025 State of AI slide deck.
The slide decks from 2018 - 2024.
The most recent 145 posts from Nateās Substack.
So thatās about 1,200 slides, but only about 50k words. Much of Nateās work was paywalled and the Inoreader preservation method I employed didnāt get those. I will have to circle back to this, because I ended up making some planned changes to Parabeagle. There are a couple people whoāve given it a try, but the ability to pass finished Chroma collections to others has just been a theory.
Once that passing stuff is tested, I have to get the rest of his content, which is going to probably involve Playwright. I could just manually plow through the missing articles, but this is one of those tools where I feel the need to level up. And then thereās the matter of the YouTube transcripts.
The transcript format is terrible ā¦ but if fed to ChatGPT ā¦ well, take a look at SelfCare Saturday: Avoiding AI Psychosis. That was the first instance where I wrote something, then included some slop at the end. If you look below the conclusion in this piece, youāll see a Slop Coda. I think thatās going to be a regular thing now - when thereās an AI contribution to what Iām writing, Iāll make that available, too.
Big Picture:
I am widely read, having never owned a TV in my adult life. Generally, when something comes up, I will have at least done a bit of reading on it. The combination of AI, environment, geopolitics, and the windup to the Trump succession fight have got me feeling a bit overloaded.
Iām trying to find balance and I donāt think I can be coherent in expressing that, so Iāll resort to a bullet list.
Video news is keeping me abreast of things; I can consume that while doing other things, pausing when something truly important surfaces.
I worry I am not investing enough time in going deep into things, but then I look at the AI facet; itās in everything I do, Iām using it to get abreast of changes.
If I am the only one benefitting, I am NOT benefitting, Iām failing my fellow travelers. That would be the two commercial things Iām doing, the folks who continue lurking Signal despite the general breakdown that happened a year ago, and ā¦ you, constant reader.
The Cicada 3301 people are the light of my life any more, a busy, happy corner of my online existence. It pains me that Iām not doing the detail work needed to move Shall We Play A Game? forward in the fashion it deserves.
Conclusion:
I really like what Iām doing now, the improvements to my health last summer have mostly stuck. I still have to be careful about what I eat and when that happens, but Iām mentally present and focused most all the time any more.
A big part of the liking that I donāt think Iāve mentioned is the nature of the business - having escaped the conflict-centric stuff that dominated my life for fifteen years, I am helping the people around me, as we build something that will help our as yet theoretical customer base.
My duties with Cicada 3301, combined with the dead ends that keep coming around because they didnāt read my retirement notice, means that I do have to stay on top of security related stuff. But thatās also nice, after so many years of bathing in hate, while filled with harmful intent. I felt competent in that field, which was good, but I did not like the person I had become while doing so.
Slop Coda:
Original ChatGPT conversation. The summary isnāt terrible, but there are some obvious whoopsies if you actually sit and listen to the video. The bit about the two million liters of water in particular I feel does a disservice to what was communicated in that segment.
š§ State of AI Report 2025 ā Executive Summary in Plain English
The State of AI Report 2025, released by Nathan Benaich and Air Street Capital, distills eight years of tracking the global AI ecosystem.
This yearās message is blunt: the model IQ race is over ā the infrastructure wars have begun.
Across the 313 slides, three compounding forces define the new era:
The capability-to-cost curve ā how fast useful intelligence gets cheaper.
Distribution ā who controls the interfaces that deliver that intelligence.
Physical infrastructure ā the real-world limits of power, siting, and capital.
1 Ā· Capability-to-Cost: Intelligence per Dollar
AI capability per unit cost is improving faster than anyone planned.
Independent trackers such as Artificial Analysis and LM Arena (cited in the report) find that effective capability doubles every three to eight months ā roughly 3ā7Ć faster than Mooreās Law.
Googleās models lead at about 3.4 months; OpenAI follows at 5.8.
Example: GPT-5ās 400 k-token input costs are ā 12 Ć cheaper than Claudeās and ā 24 Ć cheaper than GPT-4.1ās ā resetting product economics every few months.
Because quality differences are narrowing, routing intelligence ā deciding when to call a cheap model versus a frontier one ā has become the real differentiator.
At scale, AI systems now process ā one quadrillion tokens per month; even a single basis-point gain saves millions.
Funding data also reveal that model launches trail fund-raising rounds by 50ā80 days, implying many ābreakthroughsā double as financing signals.
Bottom line: capability keeps rising while cost collapses. Margin now depends on architecture and routing, not just model IQ.
2 Ā· Distribution: The Rise of Answer Engines
The report frames the browser as the new operating system for AI.
While it doesnāt name products, its description fits ChatGPT Search, Perplexity, and other āanswer engines.ā
External data show these already rival early-Google traffic.
For businesses, this means Answer Engine Optimization (AEO) replaces classic SEO. To stay visible, content needs structured data, API endpoints, and citation-friendly formats that LLMs can parse.
Economically, referral conversions from AI answers (~ 11 %) now match paid-search performance.
Yet most answer engines still depend on Googleās web index ā creating a paradox: Google supplies the corpus; OpenAI-style products capture the intent.
3 Ā· Infrastructure: Power and Permits as the Bottleneck
AI scaling has hit the wall of physics.
Projects like Stargate ā a proposed 10 GW, $500 B training cluster ā symbolize the constraint.
Each 1 GW data-center complex costs ā $50 B to build and ā $11 B per year to run, drawing the electricity of a mid-size city.
The U.S. alone faces a 68 GW power shortfall by 2028.
Local opposition and water use worsen it: a 100 MW site consumes ā two million liters of water per day.
Site selection, cooling, and energy sourcing have become geopolitical issues.
Scaling AI is now a matter of atoms as much as bits.
4 Ā· Reasoning Gains ā and Their Limits
The āreasoning raceā dominated 2025: OpenAI (o1 ā GPT-5), DeepSeek (R1), Anthropic (Claude 3.7 ā 4.5), and Google (Gemini 2.5).
But the deckās āIllusion of Reasoning Gainsā slides show many improvements within noise bands that vanish under stricter tests.
Some systems exhibit ātest awarenessā ā performing better only when they detect evaluation (the AI Hawthorne effect).
Reasoning is improving, but benchmarks often overstate it.
5 Ā· Open vs Closed Models
Slides 42ā46 make the trend clear: Chinaās open-weight ecosystem (DeepSeek, Qwen, Kimi) now dominates fine-tuning and deployment, while U.S. frontier labs (OpenAI, Anthropic, Google) keep raw performance leadership.
OpenAIās gpt-oss-120B and -20B are its first open models since GPT-2 ā a partial shift toward openness.
Reality is a spectrum: closed models for high-stakes tasks, open models for regulated or local use, and smart routing between them.
6 Ā· Sovereign AI and Global Power Politics
National āsovereign AIā projects are everywhere, but most still run on U.S. clouds, NVIDIA GPUs, and Western software.
True independence is rare ā these deals are often data-center siting agreements, not technological autonomy.
7 Ā· What the Talk Skipped
Three major sections in the 313-slide deck did not appear in the talk:
AI Safety & Alignment: Chain-of-Thought monitoring, transparency vs capability trade-offs, and the emerging monitorability tax.
Science & Industry Applications: advances in chemistry, biology, materials science, and robotics (Slides 54ā63).
Survey & Predictions: data from 1,200 practitioners and Air Streetās forecast accuracy review.
These complete the reportās six dimensions: Research, Industry, Politics, Safety, Survey, Predictions.
8 Ā· Implications
For builders ā Expect capability-per-dollar to double every 4ā6 months. Your edge is routing and AEO, not bigger models.
For investors ā Watch who controls compute, distribution, and energy permits. NVIDIA-cloud dependency loops remain.
For individuals ā The cost of intelligence is collapsing; leverage comes from intent and workflow design, not access.
Closing Note
This summary faithfully represents the first third of the official report ā the Research and Industry sections ā and conveys its strategic message with only minor interpretation.
Added clarifications bring it into full alignment with the State of AI Report 2025 slides (Slides 7ā9 & 42ā46, Air Street Capital).