Tap to play. It never plays by itself, and the captions are in the picture.
DeepSeek shares its engine as Washington shows it out. DeepSeek open-sourced DeepInfer, the engine behind its own services, under Apache 2.0. Meanwhile, a Homeland Security order bars DeepSeek tech from US civilian agency systems, CNBC reports. Meta paused Muse share links after researchers found 6,800 of them in passive DNS records. A Linux Foundation survey says AI cuts code authoring time 52 percent, but maintainers estimate 25 percent of pull requests are slop. And Salesforce agreed to buy Braze for 5.66 billion dollars. The Linux Foundation's advice: write down AI rules, and require disclosure and testing.
Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics
DeepInfer's GitHub page says it powers DeepSeek's production services and supports DeepSeek-V4.1-Flash on Nvidia SM100 GPUs and Huawei Ascend 950 NPUs under Apache 2.0. A KV cache system is not included.
Why it matters: Teams serving open models get DeepSeek's own serving stack, with deterministic sampling and structured outputs, but not a complete service out of the box.
CNBC reports the order, issued July 24 and made public Monday, makes civilian agencies remove DeepSeek tech, including third-party products built on it, and bars buying it. It excludes the Defense Department.
Why it matters: The order reaches vendors whose products are built on DeepSeek models, not just DeepSeek's own app; agencies can request exceptions, CNBC says.
Ars reports security firm Claroc found the links through passive DNS because each URL carries the chat's title, exposing a loan application and a revenue review. Meta says it paused sharing while it reviews.
Why it matters: Muse needs access to most of a user's accounts to act for them, so one URL design choice turned private work chats into searchable pages.
LF Research says respondents estimate a 52% cut in code authoring time, while maintainers estimate 25% of incoming PRs are low-quality AI slop, 31% in large projects. Only 35% of projects have a formal GenAI policy.
Why it matters: For teams using or maintaining open source, LF points to written AI policies, disclosure and testing; 21% of respondents say complex tasks take longer with GenAI.
CNBC reports Braze holders get $47.50 cash per share, versus Monday's $30.04 close. The deal, expected to close in fiscal 2028, follows Salesforce's planned $3.6 billion Fin deal.
Why it matters: Salesforce keeps buying AI customer tools, after Fin, to build out its agentic offerings for enterprises, CNBC notes.
Cohere says its 7B model cut calibration error on BigMathDigits from 0.059 to 0.028, a 52.5% drop versus the strongest single-confidence baseline, and by 26.8% on GPQA and 36.2% on SimpleQA.
Why it matters: Reliable confidence is what lets an agent send doubtful answers to a person; Cohere says it has not yet tested using it to verify or abstain.
Anthropic says an internal research model worked mostly autonomously for 64 hours with four follow-ups, generating 318,484 designs. Two labs tested 211; 17 met its binder criteria. It does not claim an antivenom.
Why it matters: A detailed case of agent-run science with stated limits: humans steered, outside labs checked, and no multi-toxin design passed for more than one toxin.
On the Frontier watch →Databricks says Wine.com launched production email campaigns 23 days after kickoff and syncs 100 audiences daily, and its own Summit campaign for 30,000 attendees got a 10 point conversion lift over a control group.
Why it matters: These are customer quotes in a vendor post, not an audit, but they show marketing agents working on governed data, with people setting goals and guardrails.
TechCrunch reports Figma named 20 new MCP partners, including Linear, Notion and Mistral, at Config in India. Field said MCP is not a perfect protocol and that exposing more data raises the chance of abuse.
Why it matters: Field's comments suggest tool vendors will adopt MCP while limiting what data they expose to agents.
Researchers report that on invalid-proof cases Nemotron and Llama requested protected actions in 52.1% and 54.2% of trials, yet a deterministic tool gate committed none. A valid proof from a compromised signer still passed.
Why it matters: Check what agents execute, not only what they ask for: enforcement at the tool held, but the authors note proof validity is not factual truth.
On the Frontier watch →OR-License draws 31 instances from paid operations research engagements totaling USD 3.68 million. Across 11 LLMs, the best Claude Code setup reached 88.39% strict feasibility and a DAQ of 0.87 against experts' 1.
Why it matters: Several models beat experts on objective quality when they delivered, but none matched experts once infeasible work scored zero, the authors report.
TechCrunch reports Yelp's Check for Me agent calls a restaurant for the user and texts back the wait. Yelp also added OpenTable booking, chat memory and a $30-a-month Smart Lead Reply agent for businesses.
Why it matters: Consumer voice agents are calling businesses, and TechCrunch notes a user's agent may end up talking to the restaurant's agent.
A person familiar told CNBC the deal could be worth around $45 billion and would add a second 460-megawatt data center at the same West Virginia site as the first. Anthropic declined to comment.
Why it matters: It adds detail to the earlier report as rivals CoreWeave and Nebius fell 28% and 27% in a month on debt worries, CNBC says.
CNBC reports Arda builds an agent for manufacturing engineers and has deployed its models with Anduril. Altimeter, Accel and Khosla led the round; McGrew says robotics is next.
Why it matters: McGrew says updating automation for a new design takes three to six months; cutting that could make fast-changing products buildable in the US.
Reviewing Anthropic's August Risk Report, METR says its methods might not catch higher risk and covert-capability evals poorly proxy evading monitoring. Anthropic could redact; it made no redactions.
Why it matters: METR urges end-to-end red-teaming of monitoring and says private examples made it think Anthropic's monitoring is weaker than public evidence suggests.
openai/codex, Apache-2.0 and created in April 2025, gained 496 stars since our last check, the largest gain among the repos we track this run.
Why it matters: Terminal coding agents keep drawing developers; Anthropic's claude-code sits at 152890 stars.
feder-cr/git-ai, created October 15, gained 425 stars since our last check and has grown from 2367 stars in our October 19 morning Drop. No license is listed.
Why it matters: Version control built for agents is drawing fast interest, though the missing license limits how teams can reuse the code.
Leonxlnx/earshot, MIT-licensed, CPU-only and created October 15, gained 243 stars since our last check, up from 1583 in this morning's Drop.
Why it matters: Audio remains a gap for most coding agents, and this CPU-only open tool keeps drawing interest.
A Peec AI researcher writes that a site of 50,000 invented statistics, built for about $50, was dropped by Google but drew about 72,000 ChatGPT-User requests in 30 days and citations in a paper and bank research.
Why it matters: Agents that ground answers on fetched pages can launder bad numbers; the author says rerankers score relevance, not truth.