- AI Fire
- Posts
- 👀 a16z Flags AI’s Next Big Test
Sponsored by
Nvidia’s chips are already in high demand while paid consumer AI adoption is still tiny. a16z’s latest report suggests the boom could have much further to go, with everyday use holding the next big opportunity.
What's on FIRE 🔥
IN PARTNERSHIP WITH SKYDIVE
What would you delegate if you had a whole team?
With Skydive, you can build a team of AI agents to take work off your plate.
Create agents for customer support, sales, marketing, engineering, ops, and more. Give each one a role, connect the tools they need, and hand off the work.
Your agents can work on their own or together, sharing context and handing off tasks to get bigger jobs done.
Start with one agent. Build a whole team around you.
AI INSIGHTS
AI money is pouring into infrastructure much faster than consumers are paying for AI. According to a16z, every $100 invested in the AI supply chain roughly breaks down like this:
$50 goes to chips
$20 to power
$15 to networking
$15 to cooling, buildings, and land
Meanwhile, paid consumer adoption is still tiny. By April 2026, only 2.2% of U.S. households had an AI subscription, with subscribers paying about $31 per month on average.
Enterprise adoption looks broader, with 69% of S&P 500 companies mentioning AI deployments. But a16z says many still do not track clear, AI-specific business outcomes. That creates a strange market: enormous spending on GPUs, energy, data centers, and networking while actual AI usage is still early.
Key takeaway: The AI boom is currently being built from the infrastructure layer upward. The next major growth phase depends on turning all that compute into workflows people and businesses use every day.
PRESENTED BY ATTIO
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
It’s your always-on revenue engine: agents and workflows build pipeline, chase every buying signal, and move deals forward alongside your team.
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?
AI SOURCES FROM AI FIRE
1. Free Guide: I Tested GPT-6.1 Sol Inside Codex on 3 Hard Tasks. Can THIS Beat Claude Sonnet 5.5? GPT-6.1 Sol looks powerful yet really cheap. I wanted to see what happens when you actually put it to work, so I tested it across 3 hard tasks and compared the workflow and results with Sonnet 5.5.
2. 5 Free AI Agent Skills Tested for Research, Coding, Video, Security and Better UI. You’ll see the installation process, how each one works, real use cases, and ready-to-use prompts for putting them to work yourself. All available right now.
TODAY IN AI
AI HIGHLIGHTS
🔌 Meta just launched Muse Gadgets, letting developers build Muse-powered devices with Raspberry Pi or ESP32 boards. Nat Friedman also announced 5,000 free Home Links for Muse subscribers, while supplies last.
👨💻 Anthropic just added mods to Claude Code, letting you change how it works or looks using JavaScript or TypeScript. You can even describe the add-on you want & ask Claude Code to build it.
🕵️ OpenAI says it shut down a campaign involving 15,000+ accounts that tried to extract its models’ hidden reasoning. It linked a core group to people associated with Moonshot AI, the company behind Kimi.
👀 Tavus previewed Griffin, a lifelike avatar that can see, hear & react during live video calls. In the company’s test, 48% of participants thought Griffin-Lite was human. Access is limited to trusted testers for now.
📜 Carter Church says OpenAI’s GPT-6 Astra cracked a 217-year-old coded letter to one of Napoleon’s generals in about 6 hours. He started with one image & one prompt, then Astra handled the research and codebreaking.
🎙️ Microsoft launched MAI-Transcribe-2-Streaming, turning speech into live text in 60 languages & ranking No. 1 on Artificial Analysis for accuracy. MAI-Voice-2.1 and its faster Flash version also arrived, supporting voice replies in 23 languages.
💰 AI INFRASTRUCTURE FINANCING: Broadcom agreed to lend Anthropic up to $42B to lease AI chips developed with Google. The deal could cover about one-third of Anthropic’s $125.2B five-year chip lease commitment, with debt that could later convert into Anthropic shares.
HOT PAPERS OF THE WEEK
1/ MaLiang-Harness lets AI build and fix images with code
Researchers from National University of Singapore, Fudan University, and Tencent, including Shuicheng Yan, introduce MaLiang-Harness, a system where multimodal models create images and videos through executable code, inspect the result, then revise it when something looks wrong. GPT-6 Astra achieved 100% generation success, with 96% of image tasks meeting all quality thresholds. What it means: AI-generated visuals could become much easier to control, edit, and debug than today’s one-shot image generators.
2/ Researchers map how robots can learn new tasks without retraining
Researchers from Knowin AI, Tsinghua University, CUHK, HKUST(GZ), and other institutions review how robots use in-context learning to adapt from demonstrations, corrections, and past experience without changing model weights. The paper groups current methods into four main approaches, from context-conditioned policies to world models and agent-based execution. Big idea: Future robots may learn new tasks more like humans, by watching, trying, remembering, and improving from experience.
3/ Raven builds an “AI manager” for teams of specialized agents
EverMind AI introduces Raven, an open-source system that can combine agents such as Claude Code, Codex, OpenClaw, and Raven’s own specialists into one coordinated workflow. Raven breaks a complex goal into subtasks, assigns each one to the right agent, tracks dependencies, and reuses past experience through memory and skills. What it means: Instead of relying on one super-agent, complex work could be handled by a coordinated team of specialized AI agents that improve over time.
NEW EMPOWERED AI TOOLS
🧑🏫 Gauth Unlimited Digital Canvas turns AI lessons into an infinite whiteboard where concepts, notes, formulas, and tutor answers unfold visually in one space.
🛒 Veltrix AI for E-commerce analyzes sales, costs, customers, and marketing to find ways to improve profit and cut unnecessary spending.
📚 Teachoo helps students solve homework with guided questions, hints, feedback, and step-by-step explanations from text or photos.
🎬 Clueso MCP lets Claude, ChatGPT, and other AI agents create and edit videos from ideas, decks, recordings, or references, with everything staying editable.
⚡ Jev turns unstructured data into fast, structured AI decisions with probabilities your software can act on directly.
AI BREAKTHROUGH
Two new open-source releases are making AI model training more accessible for smaller teams.
Pleias and the AI Alliance released SYNTH, a synthetic training dataset built by expanding 58,698 Wikipedia articles into roughly 75 billion tokens across 8 languages.
SYNTH combines pre-training, reasoning, and instruction-style data into one pipeline.
Pleias says models trained on it can maintain strong factual precision with 10–140× fewer training tokens than comparable web-trained models.
The experiments include the tiny 56M-parameter Monad and larger Baguettotron models.
At the infrastructure layer, Ai2 released Olmo-core 3, a new open stack for training large MoE models. By keeping experts resident on GPUs instead of repeatedly moving weights, Ai2 measured about 2.7× higher throughput in one 47B-parameter test.
→ Better synthetic data and more efficient training infrastructure could let smaller research teams experiment with capable language models without needing frontier-lab-scale datasets or compute.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply