A continuous newsletter on enterprise AI, RAG architecture, agentic systems, and the changing GenAI landscape — by Priyanshu Shekhar Sinha.
Apple opened the Foundation Models framework to third-party developers, shipped a 7B on-device model that runs on every iPhone back to the 16 Pro, and quietly named Anthropic — not OpenAI — as the cloud partner for th...
Fable 5 shipped last Tuesday with a fully generative dialogue system across roughly 1,400 named NPCs, running a fine-tuned 8B model on-device on Series X and a streamed inference path on Series S. It is the first AAA ...
An internal OpenAI model disproved Paul Erdős's 1946 unit-distance conjecture autonomously, using infinite class field towers from algebraic number theory. Tim Gowers called it 'a milestone in AI mathematics.' This is...
Andrej Karpathy announced last week that he's joined Anthropic's pretraining team under Nick Joseph, working on using Claude to accelerate pre-training research itself. The headline is the name. The story is the mandate.
Trump cancelled the signing of an AI executive order this week, telling the pool reporters that he was worried it would 'undermine America's competitive edge'. The absence of a signature is, in this case, more reveali...
A Miami startup nobody had heard of three weeks ago shipped a 12M-token context window on a sparse subquadratic attention architecture, claims 52x speedup over FlashAttention at 1M tokens, and priced it at roughly a f...
Pope Leo XIV released his first encyclical yesterday — 235 pages on safeguarding the human person in the age of AI — and broke a 2,000-year tradition by personally presenting it. The man standing beside him at the lec...
Everyone wrote about Gemini Spark, Antigravity, Omni. Almost nobody wrote about the pricing slide. Google quietly cut AI Ultra from $250/mo to $100/mo at I/O. That's a 60% cut on the top tier in the middle of a subscr...
Zhipu trained GLM-4.7 on 100,000 Huawei Ascend 910B processors. No H100s, no H200s, no Blackwell. The technical merit of the model matters less than the proof of concept: China can now train a frontier-adjacent model ...
$30B round closing this week at a $900B+ pre-money, Q2 revenue tracking to $10.9B, and the first time Anthropic has been worth more than OpenAI on paper. The number people will quote is $900B. The number that actually...
Notes from yesterday's I/O keynote — Gemini 3.5, Gemini Omni, Antigravity, Spark, Universal Cart. The story isn't the model numbers. It's that the entire product surface pivoted to agentic in one keynote.
A practical migration note for teams running on 2.5 Flash. What's measurably better in 3.5, what's worse (price), the four config changes you'll have to make, and a checklist for cutting over without breaking your fun...
Six things that earned a tab and a sticky note this week — Anthropic's compute deal, CAISI's pre-launch evaluations, the open-weights frontier closing in, and one quiet shift in the agent stack.
Anthropic shipped Opus 4.7 yesterday and announced a SpaceX compute partnership: 220K+ NVIDIA GPUs at Colossus 1, doubled Claude Code rate limits, and exploratory work on orbital data centres.
The U.S. Center for AI Standards and Innovation will now evaluate frontier models from Google DeepMind, Microsoft, and xAI before they're publicly available. The first concrete operationalisation of pre-deployment gov...
DeepSeek shipped V4-Pro (1.6T MoE) and V4-Flash (284B MoE) on April 24 under MIT licence. V4-Pro is now the #2 open-weights model on the Artificial Analysis Intelligence Index, behind only Kimi K2.6.
The model OpenAI confirmed for an April 14 launch — internal codename 'Spud' — finally shipped on April 23. Branded GPT-5.5, not GPT-6. The renaming is the most honest thing about the release.
Bloomberg reported today on Google's new agent line. The strategic story is more interesting than the demo. Here's what I think the bet is.
Model Context Protocol's adoption numbers are extraordinary — 970× growth in 18 months, 9,400+ public servers, 78% of enterprise teams running agents in production. A status check from a working integrator.
Meta shipped Llama 4 Scout (10M context, MoE) and Maverick (1M context, 128 experts) earlier this month. The context-length number is real. The 'do we still need RAG' question still has the same answer.
Anthropic previewed Claude Mythos last week and didn't ship it. The model can find zero-days at scale. The decision to gate it behind Project Glasswing is the most interesting deployment story of the year so far.
Anthropic's run rate quietly passed OpenAI's, and most of the delta is enterprise agentic workflows — not consumer chat. A short field report on what's actually driving that.
Mistral Large 3, Gemma 4 (Apache 2.0), and DeepSeek V3.2 all shipped in the past few weeks. The open-weights frontier is no longer a year behind — it's a quarter behind, sometimes weeks.
Field notes from the GenAI frontier — what to expect from this newsletter, who it's for, and how often it'll land.