Research archive
A dated record of AI releases, prices, policy changes, and San Diego events James was watching. This is an archive, not a live feed. Last snapshot: .
How leading models compared when this snapshot was published.
The top three labs now sit within 12 ELO points (head-to-head preference ratings) of each other — close enough to be statistical noise. For a buyer, that tie means you choose on latency, cost, and tooling fit, not arena rank.
What launched and why it could matter to a business.
Opus 4.7's 1M-token context window leaves beta and ships to everyone. The consequence for active engagements is concrete: a full-codebase RAG index (the retrieval layer you build to feed a model only the relevant files) can be retired in favor of a single-prompt context load on the 200K+ files-per-prompt path. The variable that now decides your bill is cache hit rate.
Anthropic's frontier reasoning model gets a 1M-token context window in beta. The payoff is brute force: drop a medium-sized monorepo (one big repository holding many projects) into a single prompt and analyze the whole codebase at once. That window carries a pricing premium over the 200K-window tier, which moves cache hit rate (how often the model reuses text it already digested instead of paying to process it again) from clever optimization to budget line. It is load-bearing for cost.
Software that changed the cost or effort of putting AI to work.
MLX 3.0 brings unified-memory model serving to Apple Silicon, collapsing the CPU/GPU transfer overhead (the copy tax of shuttling data between processors) for on-device inference. For SoCal teams running edge-AI prototypes on M-series workstations, it bends the local-development cost curve — and may pull some 'GPU-required' workflows back to laptop-class hardware.
Vercel AI SDK 5 makes streaming tool calls the default, not the opt-in. For Astro/Next production AI surfaces, that deletes a meaningful slab of glue code — the wiring you wrote by hand and then forgot you wrote. Practical read: prototypes ship a week earlier; production review cycles don't budge.
An independent crawl puts llms.txt adoption above 38% among the top-1,000 ranked sites in technical-content categories — up from under 8% when Q1 opened. llms.txt is a plain-text file that tells language models how to read your site; GEO (generative-engine optimization, tuning content so AI engines surface it) has stopped being a trailing-edge bet.
Changes from regulators and standards groups.
Public comment is open on proposed rules requiring disclosure of AI-generated content in commercial communications. It reaches any consumer-facing AI workflow; regulated-industry buyers should expect compliance review to pick this up in Q3.
Funding, acquisitions, and other shifts in the AI market.
The aggregator passes Q2's frontier-model price cuts straight through: top-tier reasoning models (the ones that work through a problem step by step) now run 18-24% cheaper per million tokens than they did in Q1. If an engagement is burning real token spend, re-bid the next 90 days.
Local gatherings for business and technical communities.
May 2026, at the Google Cloud San Diego venue: a GenAI and agents meetup. The CFP (call for proposals — pitch a talk) is open to practitioners, and sponsor slots run from venue to food. For builders, it's the densest room in San Diego right now.
Recurring San Diego meetup for GenAI, LLM, and agent practitioners. A strong room for practitioner-tier conversation and informal benchmarking (comparing notes on what's actually working in production) among local builders. Open to sponsor and speaker proposals.
Five days at the Hyatt Regency La Jolla, June 1–5: conference, workshops, bootcamp. The closest San Diego gets to a pure-ML (machine-learning) conference this cycle. Watch the sponsor and speaker calendars for 2027.
Methodology
James collected these items from public model leaderboards, product announcements, regulators, research groups, local event calendars, and industry publications. Every entry keeps its original date and source link.
AI changes quickly. Rankings, prices, products, and event details may no longer be current. The archive shows what information was available when recommendations were made.
For longer explanations, visit Writing .
RSS feed for writing essays Ask what matters for your business