DeepSeek open-sources DeepInfer 🐳, AI makes smart TVs dumb again 📺, ~$100/day AI on-call 🔭

DeepInfer, the software that runs a trained model, supports NVIDIA SM100 GPUs and Huawei Ascend 950 chips, and DeepSeek says future models' inference code will land in the DeepInfer repo. DeepSeek calls it fully open source, yet the README says a complete service needs parts this release omits, such as a KV cache system and load balancing across replicas. Anyone self-hosting V4.1-Flash can run it offline now but must supply those parts themselves.

NorForge researcher Snorre Fagerland watched the Chinese-speaking hacker's agents try more than 1,200 break-ins and compromise over 100 sites; Reuters identified victims including Deere, SharkNinja and Cummins. He estimates the campaign, apparently using DeepSeek and Claude Code, cost about $50 in tokens per successful hack, each typically taking around 23 hours. Fagerland calls his find real-life evidence that commercially available agents can pull off complex intrusions start to finish without human intervention.

A Homeland Security notice published Monday also bans civilian agencies from buying DeepSeek products; Commerce and the Navy had already banned it on government devices. DHS warns that use by government staff or vendors could send sensitive data to China, and that the models "may contain malware". The notice extends to third-party services built on DeepSeek, so vendors building on it can no longer sell those services to civilian agencies.

A judge on Monday granted prosecutors' request to keep former OpenAI engineer Abdulrahman Alfozan detained as a flight risk. Their filing says after opening the folder he asked OpenAI's own productivity agent who at the company could join a high-level meeting with ministers in Saudi Arabia. The recording charge carries up to five years, but the government expects it may bring more serious charges as evidence mounts.

The round of about $30 billion could value OpenAI at over $1.4 trillion, sources told Reuters, up from $852 billion in March. OpenAI is telling investors it has hit $50 billion in annualised revenue and expects $70 billion by year-end; it wants to close before Anthropic's IPO disclosures invite comparisons. Some investors see the rush as locking in money that might otherwise go to Anthropic's expected mid-November listing.

Legal filings show Disney is willing to pay $180 million over three years to hire 66 Character.AI staff; it has also bid $20 million for part of the startup's assets. That follows Disney hiring Character.AI chief executive Karandeep Anand last month as its first chief technology officer. Gizmodo says bankruptcy law frees Disney from liability for Character.AI's ongoing legal troubles, but the chatbots' risks to vulnerable users could become a headwind.

Projects like dumbtv and cleanyourtv have owners point an AI agent at a GitHub repo to strip ads, tracking and bloatware. One user says an afternoon with Claude Code disabled 25 bloat apps and swapped a 192 MB launcher for an 83 MB one. Wider uptake could disrupt the smart TV ad industry, which profits from selling viewing data, though one person reports bricking their Android TV with Claude.

Cloudflare found each reviewer session's first request reused the cached start of its prompt only 14% to 26% of the time, because job-specific details came before the cache marker. Fixing that order was the biggest single saving, though Cloudflare credits model and effort changes with $1.06 of the $1.67 drop. Teams running agents at volume should check first-request cache hits, and whether cheaper effort settings lose real findings, as some of Cloudflare's did.

For an attack the model was missing, engineers helped prioritise ideas and agents implemented and tested them; the winning setup averaged model snapshots from the start of a shorter training run. Stripe estimates the gain is worth tens of millions of dollars a year in fraud prevention for its merchants. An earlier jump to 11 billion training payments looked better, then fell behind as late fraud reports arrived, a warning wherever labels arrive weeks late.

One agent holding all the domain knowledge started hallucinating in production, so Meta moved that knowledge into specialist agents that each see one domain. The router handing them work follows a fixed protocol with no LLM reasoning, and config prep for a data column fell from about a day to under an hour. Any team whose agents mix up conventions across subsystems can apply the same split: deterministic routing, isolated domain reasoning.

The Neural Engine, an energy-efficient accelerator in every M-series chip, goes unused by MLX and llama.cpp. It reads memory more slowly than the GPU, so MacPaw has it check a batch of guessed tokens on each pass, and its energy lead holds even against MLX using the same guessing trick, at similar output quality. Neither unit won every workload, so apps running local models on Macs should pick the backend per request.

Since reinforcement learning pushes models toward whatever scores highest, the report argues flawed benchmarks can steer them the wrong way. Maintainers would not have merged about half the AI fixes that passed the SWE-bench Verified coding test, and GLM-5.2's SWE-Bench Pro score fell from 78.8% to 57.3% once an answer leak was closed. For buyers shortlisting models, it says even scores from the best benchmark should not be read alone.

Tramèr wants a cheap first pass that asks only whether anyone would want to read a paper, since writing one is becoming almost free. SaTML's quick triage, plus desk rejections, cut about 40% of its reviewing load, rejecting 95% of the 100-plus single-author papers. He also floats crediting researchers for how many people read and learn from their work, which he says would favour clear writing.

Congdon, who has worked on monitoring releases and config pushes since 2023, says agents can catch regressions with few false alarms by running endless queries and log scans. His prototype caught regressions in systems he had zero understanding of, and letting plain code run an agent-written monitoring plan cut noise and token use. He estimates Haiku-class models could watch a moderately large tech company's whole fleet for about $100 a day.

At the IMF meeting in Bangkok, Clark cited AI's top score on Humanity's Last Exam rising from 8% in January 2025 to 67.7%, and GPT-6 Astra finding errors in experts' answers to unpublished physics problems. He says Anthropic's researchers now write almost 10 times as much code as in the company's first four years. His ask is for countries to build lab and robotics capacity now, and to pace AI's own development with restraint.

Mercury 3 is a diffusion language model, an alternative to the one-token-at-a-time design most LLMs use. It reaches 1,068 tokens per second on NVIDIA GPUs with a 1M-token context, and lists at $0.20 per million input and $0.70 per million output tokens, half that at launch. Teams building voice, support or agent products that need low latency can test it through an OpenAI-compatible API; the benchmark claims are Inception's own.

Extraction here means reading fields such as invoice totals and dates out of messy text or images, each with a confidence score. Cloudflare reports a damage-claim pipeline that starts with Clairic and Clef finished in 19.3 seconds against 61.7 for an all-LLM version, at 49% lower cost. Teams parsing documents can try it on Workers AI, but it is a preview: longer chains of arithmetic are hard for it and forecasting is basic.

Coding agents can write the code that renders a film but cannot hear it. Earshot describes sounds in words, checks sync to picture cuts and scores a launch film against seven professional ones. In the author's blind test, one listener spotted the code-made sound effects 8 times in 15, a coin flip. MIT licensed, with an MCP server for Claude Code and other agents.

Older messages fade into broader summary lines, but the agent can zoom back into any line down to the original message. The agent runs a shell, edits files and can drive a Mac's screen, using your own Anthropic or OpenAI keys under Apache-2.0. Worth trying for anyone who wants a long-running agent on their own server. The README says it is vibe-coded and to expect bugs.