Daily AI Zine · No. 001

The Lead

Today’s biggest story

DeepSeek raises $7.4 billion, and the discount frontier gets a war chest

A regulatory filing reveals the first outside round in the history of China’s most closely watched AI lab.

DeepSeek has taken outside money for the first time. A disclosure filed Thursday by Anhui Korrun, a Shenzhen-listed travel-goods maker whose subsidiary invested through a private equity fund, shows the lab raised about 50 billion yuan, roughly $7.4 billion, at a post-money valuation above 350 billion yuan, about $52 billion. That makes DeepSeek China’s most valuable AI startup, and the filing is the first real look anyone outside the company has had at its finances.

The structure says as much as the number. Tencent contributed about 10 billion yuan and battery giant CATL about 5 billion, while founder Liang Wenfeng committed roughly 20 billion yuan of his own capital. Most outside investors bought into a limited partnership with five-year lock-ups and no voting rights. Control stays exactly where it was: with Liang.

Why it matters. DeepSeek built its position as the lab that did not need anyone’s money: open weights, cut-rate pricing, and a research-first posture that kept undercutting closed Western labs. That strategy now has a balance sheet behind it, plus state-linked backers with distribution across the Chinese economy. The pricing umbrella that OpenAI and Anthropic operate under keeps shrinking, and it is now funded to shrink further, including overseas.

Labeled speculation. Reports place DeepSeek in talks for a follow-on round valuing the lab north of $70 billion ahead of a Shanghai STAR Market listing. None of that is confirmed by the filing; treat the IPO track as directional until a prospectus exists.

Why you careEvery drop in the frontier price floor improves the economics of automated pipelines like this zine’s newsroom. When the rank-and-triage tier can run on an open-weight model at a fraction of today’s cost, the per-issue budget changes materially.

T1Caixin Global07/17 T2Crypto Briefing07/17

The Heat

What moved the industry

Databricks signs at a $188 billion valuation

Databricks signed a term sheet for a strategic round at a $188 billion valuation, led by existing investor Coatue and expected to close later this summer. The company says the capital goes to Unity AI Gateway, its Genie AI coworker, and Lakebase, its serverless Postgres for agents. TechCrunch pegs the raise at roughly $3 billion, up from a $134 billion valuation in February.

Why it mattersInvestors keep paying up for the data and governance layer beneath enterprise AI, even as scrutiny of AI returns sharpens.

T1Databricks07/16 T1TechCrunch07/17

Anthropic makes Claude free for US K-12 teachers

Claude for Teachers gives verified US K-12 educators free access to premium Claude, a library of teaching skills, and a direct connection to evidence-based curricula mapped to academic standards in all 50 states through the Chan Zuckerberg Initiative’s Learning Commons.

Why it mattersThe classroom is the next distribution fight. Anthropic now competes head-on with ChatGPT for Teachers and Gemini for a generation of daily users.

T1Anthropic07/14

03reported

Washington reportedly starts dictating frontier-model access

CNBC reports, citing two people familiar with the matter, that the Trump administration is asserting control over which companies and entities can access US labs’ most powerful models, decisions Anthropic and OpenAI previously made themselves through programs like Project Glasswing and Daybreak. The White House launched a public-private AI clearinghouse called Gold Eagle this week; an official says engagement is voluntary and release decisions rest with the companies.

Why it mattersIf gatekeeping moves from labs to government, access to the most capable models becomes a policy variable, not a product one.

T1CNBC07/17

SAP closes its Prior Labs acquisition, bets a billion on tabular AI

SAP completed its acquisition of Prior Labs, the Freiburg lab behind the TabPFN tabular foundation models, and committed more than 1 billion euros over four years to scale it into a frontier lab for structured business data. Prior Labs keeps its brand, leadership, and open-source releases as an independent entity inside SAP.

Why it mattersRows and columns run most businesses. Tabular foundation models just became a category with a billion-euro sponsor.

T1SAP News Center07/17

OpenAI ships its first hardware: a $230 keyboard for Codex

The Codex Micro, co-designed with Work Louder, is a $230 keypad for driving fleets of Codex coding agents. Light-up keys show agent status, a joystick launches workflows, and a dial sets how much reasoning an agent applies. OpenAI calls it a limited run, separate from the Jony Ive device at the center of Apple’s trade-secret lawsuit.

Why it mattersA novelty, but a telling one: managing agent fleets is becoming a physical, ambient activity with dedicated controls.

T1TechCrunch07/15

Your Radar

Tools and signals near you

Claude Code Docs

Claude Code artifacts can now call your MCP connectors live

The week 29 releases (v2.1.207 to v2.1.212) let a published artifact call MCP connectors each time it is viewed, so a shared dashboard shows live data through each viewer’s own connections, with per-viewer approval before the first call. The same week added a screen reader mode and auto-backgrounding for MCP tool calls that run past two minutes.

Why you careArtifacts stop being snapshots. A dashboard you generate once can now read live data through your own MCP servers, which quietly turns Claude Code output into a lightweight internal tool.

T1Claude Code Docs07/17

Model Context Protocol Blog

MCP goes stateless on July 28, and it is a breaking change

The 2026-07-28 revision, the largest since MCP launched, removes the initialize handshake and protocol-level sessions, so any server instance can sit behind a plain round-robin load balancer. It adds MCP Apps for server-rendered UI, a Tasks extension for long-running work, OAuth-aligned authorization, and a formal 12-month deprecation policy. Beta Python and TypeScript SDKs are already out; the final spec publishes July 28.

Why you careA breaking change with a date on it. Any server you run that assumes sticky sessions needs the stateless migration, and stateless MCP becomes much easier to host on serverless platforms like Netlify.

T1Model Context Protocol Blog05/21

Netlify Changelog

Netlify defaults new sites to Node.js 24

New sites on Netlify now build and run Netlify Functions on Node.js 24 by default. Existing sites keep whatever version they pin, via NODE_VERSION, an .nvmrc or .node-version file, or engines.node, until they opt in.

Why you careThis zine ships on Netlify. Every new site you create from now on inherits Node 24; pin the version deliberately rather than discovering the change at deploy time.

T1Netlify Changelog07/07

SoftBank Corp.

Japan’s sovereign AI gets concrete: 44 firms, one national model

Noetra Corp., backed by 44 companies and organizations including Sony, SoftBank, NEC, and Honda, began full-scale R&D on a Japan-developed multimodal foundation model for robots and physical AI, with engineers seconded from member firms, AIST, and Preferred Networks. Nvidia says the program’s AI factory will run 27,500 Rubin GPUs under METI’s FRONTia project, with model weights released in stages to domestic developers.

Why you careJapan’s sovereign-AI push now has a single named vehicle, state money, and a weights-sharing commitment. If you build AI products in Japan, this is the program whose outputs and partners you will keep running into.

T1SoftBank Corp.07/16 T1NVIDIA Newsroom07/16

Learn & Test

One lesson, one experiment

Lesson · protected slot 1 of 2

Context is an attention budget, not a bucket

As agents run longer, the lever shifts from writing the perfect prompt to curating what occupies the window. Model attention degrades as context fills, a failure mode Anthropic calls context rot, so the goal is the smallest set of high-signal tokens that still produces the behavior you want. Three techniques carry long tasks past a single window. Just-in-time retrieval: keep lightweight identifiers such as file paths and stored queries, and load content at runtime instead of pre-stuffing everything. Compaction: when a conversation nears the limit, summarize it and reinitialize a fresh window from the summary, tuning the compaction prompt for recall first and precision second. Sub-agents: hand focused work to helpers that each burn tens of thousands of tokens but return only a short, distilled result. If your agent loses the thread on hour two, one of these three is usually the fix.

T1Anthropic Engineering09/29

Experiment · protected slot 2 of 2

Prove prompt caching in one file

Prompt caching bills a repeated prefix at about a tenth of normal input price. Prove it in two calls: run the identical request twice and watch the usage fields flip.

cache_test.py · copy and run

import anthropic, pathlib

client = anthropic.Anthropic()
big = pathlib.Path("README.md").read_text()  # any file > ~4k tokens

def ask(q: str) -> None:
    r = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=300,
        system=[{"type": "text", "text": big,
                 "cache_control": {"type": "ephemeral"}}],
        messages=[{"role": "user", "content": q}],
    )
    u = r.usage
    print(f"wrote={u.cache_creation_input_tokens}  read={u.cache_read_input_tokens}")

ask("Summarize this file in one line.")
ask("List three key terms from this file.")

Call one writes the cache (wrote > 0, read = 0). Call two reads it (read > 0), at roughly a tenth of the input price and noticeably faster. If read stays zero, your prefix is below the model’s cacheable minimum (4,096 tokens on Haiku 4.5) or something inside it changes between calls, usually a timestamp or unsorted JSON.

T1Claude Cookbook08/14

Signals

Early and unconfirmed

reported

Zero RL reaches a trillion parameters

InclusionAI, Ant Group’s research arm, reports training a 1T-parameter model with reinforcement learning from verifiable rewards and no human-labeled data, the first public look at zero RL at frontier scale. The paper describes the model spontaneously developing self-verification and parallel reasoning.

T1Hugging Face Papers07/16

reported

Agents that grade their own homework

A Tsinghua-led team proposes SEED: a policy analyzes its own completed trajectories into natural-language skills, then distills their effect back into itself as a dense token-level signal. Reported gains reach 45.9 points over GRPO on ALFWorld. Part of a fast-growing July cluster of self-evolving-agent papers.

T1arXiv07/16

signal

A model called Honeycomb blinks in and out of Cursor

A model labeled Claude Honeycomb EAP appeared briefly in Cursor’s model picker on July 8, listed with a 1M context window and an extra-high effort mode, then vanished within hours. Anthropic has not commented, and no such string exists in the public API. The falsifiable tells: the Anthropic changelog, the models endpoint, and Bedrock or Vertex listings.

T2Valletta Software07/15

Three Moves Today

Do, investigate, watch

→ Do

Pin the Node version on every Netlify site you own. New sites now default to Node 24; add NODE_VERSION or engines.node so builds never shift underneath you.

→ Investigate

Run the MCP 2026-07-28 beta SDKs against any server you maintain. The stateless revision lands July 28, removes the initialize handshake, and breaks sticky-session assumptions.

→ Watch

The Gold Eagle clearinghouse. If Washington starts gating frontier-model access, availability and terms for API builders can change with little notice.