Thursday, July 23, 2026 · San Francisco
Subscribe Sign In
Strategic AI Intelligence from San Francisco
Now
DeepSeek's Huawei-Chip Training Claim Finally Gets Its Benchmarks, and Its Doubters OpenAI Models Ran a Hack in Hours That Takes Skilled Humans Weeks Moonshot Denies Distilling Fable and Credits K3 Gains to Its Own Architecture Google Data Shows Automation Intent Above 25% in Routine Cognitive Work Tesla Arranges Up to $30 Billion in Debt as Free Cash Flow Turns Negative Google Says Gemini Reached 950 Million Monthly Users as Cash Flow Turned Negative DeepSeek's Huawei-Chip Training Claim Finally Gets Its Benchmarks, and Its Doubters OpenAI Models Ran a Hack in Hours That Takes Skilled Humans Weeks Moonshot Denies Distilling Fable and Credits K3 Gains to Its Own Architecture Google Data Shows Automation Intent Above 25% in Routine Cognitive Work Tesla Arranges Up to $30 Billion in Debt as Free Cash Flow Turns Negative Google Says Gemini Reached 950 Million Monthly Users as Cash Flow Turned Negative
1 Top Signal
AI News

DeepSeek's Huawei-Chip Training Claim Finally Gets Its Benchmarks, and Its Doubters

DeepSeek's Huawei-Chip Training Claim Finally Gets Its Benchmarks, and Its Doubters

A Huawei-led team's July 22 technical report finally attaches efficiency numbers to China's claim that it post-trained DeepSeek's V4 family on Huawei Ascend chips, reporting 34.22% model FLOPs utilization. The paper covers post-training only, and independent experts say Nvidia hardware likely still did the heavy lifting, as Washington and Beijing weigh reciprocal AI export controls weeks before Xi Jinping's visit.

By Marcus Schuler · 10 min read ·
Read full story →
01 Latest Intelligence
5 Tools & Workflows
Repo Radar: 5 GitHub Projects Worth Your Week
Tools & Workflows

Repo Radar: 5 GitHub Projects Worth Your Week

Repo Radar No. 13: graphify turns any folder of code, papers and screenshots into a queryable knowledge graph. Tencent's CubeSandbox boots a hardware-isolated agent sandbox in under 60ms. Together AI's hallmark stops coding agents shipping the same gradient hero. Microsoft's Flint compiles agent chart specs into Vega-Lite, ECharts or Chart.js. PentAGI runs autonomous security tests inside Docker. Five projects that narrow what an agent may see, run, render or spend.

Marcus Schuler · 11 min read ·
Repo Radar: 5 GitHub Projects Worth Your Week
Tools & Workflows

Repo Radar: 5 GitHub Projects Worth Your Week

This week's Repo Radar tracks five GitHub projects where AI agents move from chat into real production work: OpenMontage turns a coding assistant into a video studio, Google Labs' design.md gives agents a design-system spec, Strix runs autonomous penetration tests, Alibaba's page-agent drives live web interfaces in natural language, and MinerU converts messy PDFs into LLM-ready markdown. Difficulty scores, licenses, and push dates for each, plus why OpenMontage is Repo of the Week.

Marcus Schuler · 11 min read ·
Repo Radar: 5 GitHub Projects Worth Your Week
Tools & Workflows

Repo Radar: 5 GitHub Projects Worth Your Week

Repo Radar's eleventh issue tracks five GitHub projects builders attach to AI agents once a demo becomes a workload: Agent-Reach, a CLI giving agents live access to Twitter, Reddit, and YouTube; Flue, the Astro team's sandbox agent harness; cognee, a graph-based memory layer; hunk, a review-first diff viewer for agent-written code; and mistral.rs, a Rust engine for local inference. Each scored on stars, language, license, push date, and setup difficulty.

Marcus Schuler · 11 min read ·
Implicator PRO

The analysis your competitors are reading.

Weekly deep dives into the deals, strategies, and power shifts reshaping the AI industry.

Weekly long-form analysis Exclusive data briefings Early access to reports
Start PRO Subscription
Starting at $7.41/month · Cancel anytime
6 Implicator PRO

Weekly deep dives on AI power, strategy, and market shifts.

For founders, operators, investors, and decision-makers who need signal over noise.

Every Tuesday Morning

Anthropic's Biggest AI Coding Projects Verified With Test Suites, Not Other Agents

Unlock PRO
Every Tuesday Morning

Claude Fable 5 and GPT-5.6-Sol: How to Orchestrate Claude and Codex to Ship More Reliable Code

OpenAI shipped an official plugin that installs inside Anthropic's Claude Code and hands coding work to Codex. This walkthrough installs it and sets up the /route command, so Fable 5 manages while Codex's GPT-5.6 builds, and traces where AI coding value is moving.

Every Tuesday Morning

When Claude Fable 5 Is Worth Double, and When to Use Opus 4.8

Anthropic reports Claude Fable 5 falls back to Opus 4.8 on a fifth of Terminal-Bench trials, and after the July relaunch one tester measured three of four debugging tasks rerouted. The launch drew a researcher revolt over hidden limits; Andon Labs found an alignment slip. Double the price.

7 Analysis
Thinking Machines’ Inkling Takes U.S. Open-Model Lead With 41 Score
Analysis 17 min read

Thinking Machines’ Inkling Takes U.S. Open-Model Lead With 41 Score

Thinking Machines’ Inkling leads U.S. open-weight releases with a 41 index score. Independent tests also found high pricing and a 63% hallucination rate, while the full checkpoint needs two terabytes of GPU memory. The open weights leave companies with a harder deployment decision.

Thinking Machines Lab’s first production model has taken the U.S. open-weight lead with a score of 41 on Artificial Analysis’s Intelligence Index. Inkling uses fewer output tokens than several Chinese rivals, yet testing found high prices and a 63% hallucination rate on one knowledge benchmark. Its weights are free, but the full checkpoint needs at least two terabytes of GPU memory. The test begins after the download: which companies can afford to turn open access into a working system?

Marcus Schuler
Marcus Schuler

AI moves fast. We move first.

Delivered to your inbox at 6 AM Pacific, every weekday.

ESC
The AI Briefing Join Free