One line hit Hacker News very directly this week: "Small Models Have Arrived"
The same week, Nvidia reportedly moved to acquire Hugging Face for $13B. Putting them together — capability is getting small, cheap, and ubiquitous; and when capability is no longer scarce, what the big players grab for is not "whose model is strongest," it's "who holds the structure"
AI Models & Products
"Small Models Have Arrived" (HN high score). This is the judgment I most agreed with this week. For two years everyone chased big models and benchmarks, but what's actually changing daily life is small models — running on devices, cheap, good enough. For someone working solo who cares about cost and autonomy, small models are good news: plenty of things simply don't need the most expensive cloud giant
Anthropic previewed a "Model Hardware Standard." Looks low-level, is actually the most pivotal. When a model company starts defining "how hardware interfaces with models," it doesn't want a product, it wants an entire infrastructure layer. I believe one thing firmly: whoever sets the standard captures most of the value in the end. Worth tracking long term
AI Dev Tools & Agents
"The load-bearing vocabulary of Claude" hit HN. Fun and practical — analyzing which words Claude is especially sensitive to, which are "load-bearing." My takeaway: communicating with AI, word choice is itself the interface. Same meaning, swap in a load-bearing word, and the result changes a lot. That's a skill I keep honing: not writing longer prompts, but using the right key words
Gemini-3.5-Transcribe shipped; DeepSeek released a new agent harness. Speech transcription and agent toolchains are iterating fast. My stance on these is consistent: prioritize what's stable, offline-capable, and self-sovereign; anything flashy but locked to one cloud, I hold back on
Expert Takes
One a16z title landed straight: "You are not a model. Don't price per token." It argues that pricing an AI product by simply passing through token cost reduces you to a reseller of the model. What you should sell is the problem you solve and the outcome you create, not how many tokens you burned. A great reminder when I price freelance work — sell value, not compute
This week a court ruled the "administrative blacklisting of Anthropic illegal," alongside sanctions against a certain AI group. What I see: the competitive arena for AI companies is no longer only technical, but legal and political. However strong a company's model, whether it can operate, whether policy can block it overnight — these are already part of the business model
VC & Markets
Nvidia reportedly acquires Hugging Face for $13B. I read the move, not the number. Hugging Face is the hub of open models; Nvidia is the landlord of compute. The landlord buying the hub too means folding "how models get distributed" into its own map. When capability gets cheap, whoever controls the "distribution position" becomes more valuable
Google Cloud launched "Cloud Run instances" for always-on personal AI agents. Practical for independent developers. Persistent, stateful personal agents now have a low-cost home. I'll actually evaluate this — it's exactly the infrastructure the "one person + a swarm of always-on agents" work style needs
My Take
This week's theme continues into a single larger line: when AI's capability gets small, cheap, and ubiquitous, competition shifts from "whose model is strongest" to "who holds a structural position"
- Capability side: the small-model era, models burrowing into devices
- Structure side: Nvidia buying the hub, Anthropic setting the hardware standard, legal battles deciding who can operate
The takeaway for myself is direct — don't live as a "reseller of the model." If what I do is just resell AI's capability, I'll depreciate along with the model. The positions worth holding are the things that become scarcer once AI is cheap: judgment, trust, cross-domain experience, owning the outcome
This echoes a16z's "don't price per token" — sell value, not compute
Action Items
- Seriously evaluate small models — many things don't need the most expensive cloud giant; cheap, offline, self-sovereign matter more
- Price on value, not tokens — passing through token cost reduces you to a model reseller
- Watch the "standard-setting" moves over the "benchmark" news — whoever sets hardware standards and protocols is drawing the map of the next decade
- Hold the positions that get scarcer once AI is cheap — judgment, trust, cross-domain experience, owning the outcome; small models won't replace these
Sources
RSS Digest: see research/digests/2026-W35.md (273 articles this week, from Hacker News, Anthropic, Meta AI, a16z, Google Cloud, and others)