October 23, 2026 · Friday

DeepSeek reportedly closes a $15B funding round

CATL and Tencent are said to lead a round of roughly RMB 100 billion, Bloomberg reports, though the lab has not yet confirmed it.

REPORTED BY @dotey · OCTOBER 22

Bloomberg reported on October 21 that DeepSeek has raised about RMB 100 billion — roughly $15 billion — in its latest round, with battery giant CATL and Tencent contributing the most. The lab itself has not officially confirmed the figure. If it holds, the raise would hand one of the world's most-watched open-weight labs a war chest for training compute, inference capacity, and talent, sharpening a market already defined by cheap and capable models. The reported investor roster leans toward hardware and infrastructure, a signal that compute, not just software, is the next battleground.

● SAFETY

OpenAI opens its monitoring system to outside testers

OpenAI invited METR and Redwood Research to pilot an embedded evaluation of its internal monitoring system, granting the two groups deep access to its infrastructure over three weeks. Their findings helped strengthen safeguards as the system expanded, and will shape how it is scaled in the future.

● FRONTIER RED TEAM

Anthropic uses Claude to formally verify code at scale

On its Frontier Red Team blog, Anthropic describes using Claude to formally verify software across large codebases. Formal methods can prove the absence of specific bugs, but historically only worked for small programs or demanded scarce expert time. AI models are changing that trade-off, and Anthropic is testing how far the approach can stretch.

● PHOTO STORY · MODEL RELEASE

Ideogram ships its first vector model

A text prompt, or any image, becomes a clean, editable SVG.

Ideogram's new vector model turns prompts into scalable vector art with smooth gradients and sharp typography.

The new model converts a prompt into scalable vector art, or turns any image into a detailed vector. It is live in Ideogram and the API, aimed at logos, print design, and illustration — output that stays crisp at any size.

● STANDARDS

Agent Skills donated to the Agentic AI Foundation

Anthropic is donating the Agent Skills standard to the same foundation that now stewards MCP. Since December 2025, more than 70 agent platforms have integrated it, including ChatGPT, Cursor, GitHub Copilot, and VS Code.

● MODEL RELEASE

LlamaIndex releases OCRia-1 document parser

OCRia-1 pushes the Pareto frontier on accuracy, speed, and cost, reaching state-of-the-art results on complex tables, charts, and documents.

● INFERENCE

SGLang v0.5.22 lands with 950 PRs

InstantTensor brings up to 10x faster weight loading, while DeepSeek-V4.1-Flash gains up to 2.3x output-token throughput on Hopper. The release drew 245 contributors, 69 of them new.

● MODEL RELEASE

Hugging Face releases Carbon-X DNA model

The 14B model is the first to bring DNA and natural language together. It beats every open language model tested on DNA benchmarks and lands within a point of GPT-6.1 Sol at more than 1,000x lower inference cost.

● PAPER

Tencent Hunyuan ships WorkForge framework

WorkForge constructs verifiable training environments for long-horizon work agents from real-world resources, generating authentic workspaces, checkable facts, and tasks.

● ANALYSIS

Mollick charts 30 months of exponential AI progress

Using his new book as a marker, Ethan Mollick tracks AI's jump from doing 4 minutes of human work in 2024 to 17 weeks today. He notes that 180 models have passed the GPT-4 level and 19 now run on phones — a reminder that capability keeps compounding on a steep curve.

A sign that you are in an insane millenarian death cult is that you are axiomatically convinced the apocalypse will happen within a few months.
— François Chollet

● AGENTS · TOOLS · INFRASTRUCTURE OCT 22 · SIGNALS

OPENAI

Decisions API hits general availability

Near-real-time classification of text and images, plus next-step decisions, without generating a full model response.

ANTHROPIC

Claude API fully available on Google Cloud

The full Claude API from a Google Cloud account, including Managed Agents, code execution, and Opus 5.5 fast mode.

REPLIT

Replit launches Monitoring Agent

An autonomous on-call engineer that investigates alerts, locates causes, verifies fixes, and deploys.

VLLM

vLLM supports Sage-A1.5 on day one

Shanghai AI Lab's 35B MoE model with 3B active parameters gets Day-0 support for local agent serving.

AWS

SGLang launches Trainium plugin beta

Built on TorchNeuron with AWS, supporting inference serving and RL rollouts on Trainium2 and Trainium3.

PERPLEXITY

Agent API adds batch processing

Up to 50,000 requests per file, each running a full agent loop with web search and fetch in batch mode.

CHALLENGE

$100K to accelerate Muse Spark

Daniel Gross launches a contest for the fastest run of Meta's Muse Spark Open, from GPU kernels to SGLang serving.

SECURITY

AI-found PHP flaw sparks patch race

PHP2Shell was independently found by AI security teams at OpenAI and Wiz, affecting WordPress and other widespread systems.

HARDWARE

Ascend 950PR specs surface

600W and 325 TFLOPS BF16 — roughly 44% of H100 and 21% of B300, with lower energy efficiency.

● SHORT TAKES OCT 22 · PRODUCTS

DEEPMIND

10m forest-type map

ForTy distinguishes natural, planted, and farmed forests at 10-meter resolution.

FASTINO

GLiDE v1.2 tops leaderboard

Ranks first on the Hugging Face decision model leaderboard.

COHERE

CulturalRiddles benchmark

Measures cultural reasoning, where LLMs still stumble despite math gains.

REPLIT

Government edition

AI tools for US and international public servants at every level.

LM STUDIO

More Flash capacity

New US capacity cuts rate limits for DeepSeek-V4.1-Flash-1017.

PERPLEXITY

Decider for Mac

pplx-decider-flash-v1.2 runs locally with no API key.

HIGGSFIELD

Content Factory

Turns any idea into a viral reel or animated ad hook.

OPENAI

Uber rolls out ChatGPT Work

Agentic capabilities across the whole company.

VERCEL

Instant agent deploys

tmp.now removes friction, even signup, for agents to ship software.

FUNDING

Modal closes $750M Series D

GPU market and sandbox infrastructure trends discussed.

● THE NUMBERS OCT 22 · BENCHMARKS & COMPUTE

BENCHMARKS

ROAD MOPD tames teacher conflicts

● NOTES & OPINION OCT 22 · COMMENTARY

OPEN MODELS

Open models and the cyber defense balance

Nathan Lambert argues open weights will inevitably be used in attacks, but incidents will occur across every AI path — the point is to let open models help defense.

BIOLOGY

HF CEO: bio AI guardrails are too shallow

Clement Delangue says superficial guardrails block scientists and public-interest groups while being easy for bad actors to bypass.

ROBOTAXI

Automation is more than a capability threshold

Fainting passengers, medical emergencies, and vomiting show that full robotaxi operations need extensive real-world process design.

OPENNESS

The closed-lab pitch and paternalism

Strip away the safety language, the argument goes, and a handful of companies decide what users may ask and build.

SPEECH

Checking TTS pronunciation via phonemes

Decoding TTS audio with Kaldi and comparing phoneme sequences, not characters, lets you catch misread polyphonic and rare words at scale.

LEARNING

AI-generated 5-minute learning videos

Arav Srinivas argues access to experts is no longer a bottleneck, demoing a clip on how a fab makes chips.

● IN BRIEF OCT 22 · PRODUCTS

PIKA

Pika inside ChatGPT

Creators generate full videos from a single ChatGPT prompt.

LM STUDIO

Read Aloud with local TTS

ROBOTICS

Microduck preorders pass 20,000

The $399 open-source biped rode a keynote bump toward 2026's top seller.

GROK

Grok Bot integrates with 1Password

Authorized logins let the bot sign in on your behalf.

GROK

Grok Bot enters X Chat

Can reply only in the desktop or mobile app to avoid clutter.

SYNTHESIA

Interactive GTC avatar

A cartoon Jensen recommends sessions on the spot.

VIDU

Q4 Preview supports 4K

Output from 540p to 4K with 10-bit color, plus 30% off credits.