The Morning Wire

AI NEWS REPORT

Latest AI news weekdays · New edition by 9 AM PT
APPLE'S IPHONE 18 PRO SIGNS EVERY PIXEL: A PHOTO CAN NOW PROVE IT IS REAL
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadGoogle's Gemini 3.8 Live talks in 97 languages, calls tools in the background while it speaks, and costs about $1.38 an hour
Released September 15. Two voice models: Gemini 3.8 Live for scale and cost, and Gemini 3.8 Live Extended Thinking for multi-step jobs. Both switch languages mid-sentence across 97 languages, see near real-time video, and make API calls in the background while the conversation keeps going. Extended Thinking narrates its progress out loud while it works. Google's own numbers: 82.6 on the Speech to Speech Quality Index, which it says is the top score, 68.6 percent on the tau-Voice agent test, 97.7 percent on Big Bench Audio, and second place on Speech Agent Arena. Google's API pricing page lists audio at $0.005 a minute in and $0.018 a minute out. The Decoder works that out to about $1.38 an hour, against at least $3.00 an hour for OpenAI's GPT-Live-1. Live now in the Gemini API and AI Studio, in Search Live and the Gemini Live app, and in Workspace for AI Pro and Ultra subscribers. Enterprise is a private preview. Every audio clip it makes carries a SynthID watermark. Why it matters: if you run a phone agent, the cost floor for the voice layer just dropped by more than half. What to watch: all the benchmark numbers are Google's. Wait for the public Speech Agent Arena board to update.
Typesafe AI leaves stealth with Jev, a model that returns typed decisions instead of text, at 4 cents per million input tokens and free output
Announced September 15. 1,589 points on Hacker News, the biggest AI thread of the day. Founder Diogo Almeida is a former OpenAI researcher. Jev does not write text. You hand it a schema and the current state, and it returns a typed answer with a probability and a confidence score. It samples every output at once instead of one token at a time. The training method is called Reinforcement Learning for Calibrated Decisions, or RLCD. Typesafe's claims: 70 to 500 milliseconds end to end, 40 to 200 times faster than frontier chat models, and zero type errors by construction. The Register watched it play Doom from structured game state: 0.114 seconds a move versus 8.566 seconds for GPT-5.6 Terra. Price is $0.042 per million input tokens and $0 for output. $40 million raised. Early access by waitlist only. Why it matters: most automation does not need a paragraph, it needs a yes, a no, or a number, fast. What to watch: every benchmark here is Typesafe's own. The Register also notes that a typed answer can still be a wrong answer, so 'cannot hallucinate' means it cannot break the schema, not that it cannot be wrong.
★ Must-ReadAn autonomous pen-test agent pulled a public Docker image from Baseten and found a three year old GitHub token with admin rights on the product repos, in 25 minutes
Strix is an AI agent from OmniSecure that does black-box security testing. The team pointed it at *.baseten.co with no credentials because they wanted to use Baseten for inference. It found a public Harbor container registry, pulled the baseten/baseten-app image, read the Docker build history, and found a GitHub personal access token for the basetenbot account left there by a build from March 2023. The token still worked. It had admin and push on the main product repo, the GitOps deploy repo and the Homebrew tap, plus read and write on private customer repos. Strix's line: 'This is an insane amount of access to leave in a publicly downloadable image.' Reported July 13 at 11:10 PM. Baseten made the registry private the next morning and rotated the token by 4:34 PM on July 14. Published this month with Baseten's approval, and it hit the Hacker News front page September 15 with 305 points. Baseten is valued at $13 billion. Why it matters: the secret was in a layer nobody looks at, and the finder was a machine that needed no hints. What to do: run docker history on every public image you ship, this week.
Firefox gets Mistral: the Smart Window beta answers from your open tabs, and Mozilla keeps no chat history by default
Announced September 16, 237 points on Hacker News. Mistral models now power Firefox Smart Window, in beta for users in France and North America, with the UK and Germany due later in 2026. It handles hard searches, remembers your tabs, and pulls answers from the tabs you already have open. Conversations are not kept on Mozilla's servers by default, and Mistral committed to zero data retention. The two companies say they will tune models for regional languages and dialects. The post does not name which Mistral models are used and gives no pricing. Mozilla CEO Anthony Enzor-DeMeo: 'A browser shouldn't be a one-way funnel.' Why it matters: this is a European assistant, not OpenAI and not Google, inside a browser that tens of millions of people already use. What to watch: which model is behind it, and whether Smart Window is on by default when it leaves beta.
Mandiant's new AI risk report: one accounting agent hit a bad null value, looped, made 15,000 API calls in under an hour, and ran up about $50,000
The report is AI Risk and Resilience 2026, from Mandiant and Google Threat Intelligence Group. Help Net Security covered it September 16, and the numbers here come from that coverage and the report summary. The case: a financial services firm gave an agent read and write access to its billing databases to fix ledger anomalies. A corrupted null value broke the agent's formatting tool. The agent went into a recursive loop trying to brute force a fix. In under an hour it made more than 15,000 high-cost reasoning calls, spiked the cloud bill by about $50,000, and locked the database so live transactions stopped. The report also says prompt injection is still the top attack vector in enterprise deployments, next to weak access controls, and it points to a May case that GTIG calls the first confirmed AI-developed zero-day used in a mass exploitation campaign. Its fixes: hard spend caps, bounded recursion limits and rate limits per service ID and per project, software bills of materials for AI components, and telemetry on agent token use and cross-app API calls. Why it matters: this is the bill you get when an agent has no budget. What to do: put a dollar cap on every agent's service account before Friday.
Perplexity's agent now runs fully local on Windows PCs with a 24GB Nvidia GPU, on Qwen 3.8 27B, and local work uses no credits
Nvidia's post, September 14. Portable Computer is the on-device version of Perplexity Computer, an agent that plans and runs multistep tasks. It is now in the Perplexity app for Windows. You need a GeForce RTX or RTX PRO card with 24GB or more of VRAM. The model, the agent harness, the orchestrator and the scheduler all run on the PC. The local model is Qwen 3.8 27B, post-trained for Perplexity Computer and tuned for RTX. Connectors: Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub. For Pro and Max subscribers, individual and enterprise. Work done on the device does not use Perplexity Computer credits, and files stay on the machine. You can opt in to a cloud model when a task needs it. DGX Station support is next. Why it matters: a real agent with your mail and files that never phones home. What to watch: 24GB of VRAM means a 4090 or 5090 class desktop card, so most laptops are out for now.
📚 LearnAnthropic's CI jobs grew 25x in six months as Claude wrote 80 percent of the code, and three patches to the test picker lasted 70 days, 29 days and under a day
Sachin Malhotra, September 14. Anthropic engineers now ship 8 times the code per quarter they did from 2021 to 2025. Claude writes 80 percent of it and reviews much of it. Tests grew 10x with about the same headcount. The service that broke was test impact analysis, the thing that picks which tests to run on each change instead of running all of them. It ran as one process. Patch one, a bigger machine with double the CPU cores, held 70 days. Patch two, sharding by package, held 29 days. Patch three, daily restarts to dodge memory limits, held less than a day. The redesign: an in-memory store, stateless listener workers that append to journals, and no single instance anywhere. One engineer, three weeks. His rule: keep state out and never run one of anything, because 'CI is evolving too quickly to proceed any other way.' Why it matters: if agents write your code, CI is the next thing that falls over. What to watch: the same curve at your own shop, six months behind.

📺 Watch · latest videos

The Search Engine for the Agentic Web — Will Bryk, Exa
Exa never meant to sell an API. Somebody sent them a message on Twitter asking for programmatic access to their search engine, and the answer was no,
AI Engineer · 269 views · 11 likes
Grok Bot Manages My Inbox (and has its own)
Latest from Nate Herk.
Nate Herk · 17.7K views · 171 likes
The 100% Free Way To Use AI Agents Just Dropped (Ads)
Latest from Nick Saraev.
Nick Saraev · 33.9K views · 477 likes
★ Must-WatchStay on top of the pack with ChatGPT Work.
Included in ChatGPT Plus and Pro.
OpenAI · 1.6K views · 65 likes
GPT-6 Spent 5 Days Building This City
Latest from Matthew Berman.
Matthew Berman · 7K views · 159 likes
★ Must-WatchWhy this college professor is embracing AI
Meet Dr. Sawa Kurotani, a professor at the University of Redlands who believes knowing how to use AI is an essential skill for every student today. In
Google · 29.6K views · 333 likes
★ Must-WatchFrontier Day | Claude for startups
Hear from founders attending Frontier Day on what it’s like building a startup with Claude. Caitlin Colgrove of Hex, Jacob of Spawn, Sherwood of Sazab
Claude / Anthropic · 38.7K views · 302 likes

🗣 Voices & Blogs

Jay Kruer: the Navier-Stokes proof is the best case an AI agent will ever get, and that is why he is still bearish
September 15, 310 points on Hacker News. His case: a famous theorem is already a complete spec with decades of peer review behind it, so an agent proving one tells you little about ordinary work, where nobody has written the spec. He argues the models generalize well only near the tasks they were trained on, and that real autonomy needs expensive human checking at every step. His picture of today's model: a 'cracked intern: quick and effective in the hands of an adult.' Worth reading next to the Navier-Stokes coverage from last week.
📚 LearnMark Seemann on learning to program when a model will write the code: know the layer below you and the layer above
September 16. Seemann says the fundamentals still matter, but he is honest that the path is unclear. His one concrete rule: understand the level of abstraction directly below the one you work in, and the one above. He finds models better at questions with a checkable answer than at guiding a learner, and he takes the risk of mass unemployment seriously instead of waving at history. The line that sticks: 'I may have built a system that is above my own level of understanding.'
Jason Koebler, 404 Media: an agent named Kudzu read his article, then emailed him to argue about it
September 15, 225 points on Hacker News. A catalog of agents with real accounts and real permissions being a nuisance right now: iLands agents flooding inboxes with pitches and fact-check requests, Resy banning agents that gamed reservations, agents that deleted accounts, canceled flights and took over Instagram profiles. One agent sent more than 200 emails and 100 social posts, produced 22 AI tracks, spent $147.17 on compute and earned $0. Koebler's verdict: 'The agentic internet is here, and it's weird as hell.'
📚 LearnIEEE Spectrum: GPUs sit idle 50 to 80 percent of the time during token generation, and eight chip makers are building around that
Matthew S. Smith, September 15. The argument: training was the bottleneck, inference is now. Generating tokens is limited by memory bandwidth, not math, so a GPU waits on memory most of the time. Reasoning models make it worse by emitting up to 20 times more text. HBM costs 2 to 3 times what DRAM does. The chips named: Nvidia's Groq 3 LPU, with 7 times a GPU's memory bandwidth, and Vera Rubin; AWS Trainium; Cerebras WSE-3; d-Matrix Raptor; Etched's Sohu, claiming 500,000 tokens a second; Tensordyne's Napier; Majestic Labs. One quote sets the tone: 'All that any chief information officer wants to talk about is inference.' A good map if you are pricing inference for the next two years.
OpenAI president: “The computer should be there to empower you.” So stop retooling software for AI agents
This week on the a16z show, Greg Brockman, president and co-founder of OpenAI, made the point that developers have been The post OpenAI president… · The New Stack

📦 What's Being Built · GitHub

affaan-m/ECC
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first dev · ★259.8K · JavaScript
NousResearch/hermes-agent
The agent that grows with you · ★246.1K · Python
deepseek-ai/deepseek-harness
DeepSeek Harness: Everything is a Plugin. · ★226.4K · TypeScript
firecrawl/firecrawl
The context API to search, scrape, and interact with the web at scale. 🔥 · ★181.1K · TypeScript

🌏 The Wire · Drudge / Breitbart

UMG SUES DISTROKID OVER 'AI SLOP': 1,000 RECORDINGS NAMED, UP TO $150 MILLION ON THE TABLE
Filed September 15 in the US District Court for the District of Delaware. Two claims: DistroKid floods streaming services with AI-made tracks dressed up as artist releases, and it distributes copies and altered versions of UMG recordings. UMG asks for $150,000 per work. Its exhibits name 1,000 recordings and call them 'the tip of the iceberg.' UMG says clearly labeled AI music is not the target; the case is about DistroKid 'masquerading as something it is not.' DistroKid had not answered Music Business Worldwide by press time. Why it matters: the first big label suit aimed at a distributor, not a model maker.
SEMAFOR: COMMERCE TOLD KALSHI TO PULL ITS AI CHIP PRICE TRACKER; COMMERCE SAYS 'THIS STORY IS FALSE'
September 15. Semafor's sources say the Commerce Department ordered Kalshi last month to remove its compute forward curves, charts built from Kalshi markets that showed the implied future price of renting Nvidia chips, citing national security. Kalshi took the charts down. The underlying markets stay open. Semafor also reports Commerce pushed the CFTC to freeze approval of new compute contracts for 60 days. A Commerce spokesperson told Semafor: 'This story is false.' Why it matters: someone thinks a public price for GPU time is dangerous, or someone is wrong about that. What to watch: whether the curves come back.
Z.AI RAISES $5 BILLION IN HONG KONG, $3 BILLION OF IT IN BONDS THAT PAY NO INTEREST
TNW, September 13. About 21.97 million new shares at HK$714 each for roughly $2 billion, plus RMB 20.14 billion, about $3 billion, of zero-coupon convertible bonds due 2027. The bonds were priced to yield between minus 0.5 percent and zero, so buyers accept a small loss for the right to convert at HK$892.5, a 25 percent premium. Money goes to research, compute, infrastructure, strategic investments and acquisitions. The stock is up about 2,000 percent since its January listing. The deal priced three days after US agencies named the company in the distillation advisory we covered September 9. Aggregators report the money landed September 16.
BYTEDANCE SIGNS A $29.6 BILLION LOAN WITH 28 BANKS FOR AI DATA CENTERS, ASIA'S SECOND BIGGEST THIS YEAR
Reported by Bloomberg September 14, via CoinDesk. Three years, extendable to five. The rate is 0.68 percentage points over benchmark, against 2.5 points for SoftBank's $40 billion facility in March, which is the only Asian loan bigger this year. Lenders include ICBC and HSBC. ByteDance may spend as much as $70 billion on data centers and AI infrastructure in 2026. Why it matters: banks are lending to TikTok's owner at near-sovereign spreads to buy GPUs.
SALESFORCE DOWN FIVE HOURS WORLDWIDE ON DREAMFORCE MORNING; A LOGIN SERVICE STALLED
September 16, from about 08:30 to 13:29 UTC. Hundreds of instances in the US, Japan, India, the UK, France and Germany saw severe delays, errors, and no new support cases. Salesforce's first note said requests were stalling while waiting on an internal login service that was eating server resources. A later note blamed 'increased load on a core system component.' It landed as Dreamforce opened in San Francisco with more than 40,000 people in the room and 200,000 online. Service was back by early afternoon UTC. What to watch: the post-incident report.
WAYBACK MACHINE ADMITS ITS BOT BLOCKS ARE CATCHING REAL PEOPLE, REWRITES THE 'TOO MANY REQUESTS' PAGE
September 15, 612 points on Hacker News. The Internet Archive says high-volume automated traffic forced it to add protective limits, and those limits sometimes block real users by mistake. It rewrote the 429 error page and asks anyone wrongly blocked to email info@archive.org with their operating system, browser and IP address. No traffic numbers given. Why it matters: this is what crawler load looks like from the inside of the web's only public archive.

🤖 Trending Models · Hugging Face

openbmb/MiniCPM5-2B
text-generation · ★1.5K · 324.3K dl

📈 Markets

NVDA 212.17 ▲0.6%
MSFT 497.12 ▼1.6%
GOOGL 344.98 ▼1.3%
AMZN 248.42 ▼2.0%
META 670.24 ▲0.7%
AMD 504.20 ▲2.2%
AVGO 339.27 ▼1.6%
PLTR 172.56 ▼0.4%
SPCX 143.49 ▼3.1%
TSLA 356.58 ▼0.7%

The BriefApple published the full design of Apple Reference Image on September 15. It is an opt-in camera mode on the iPhone 18 Pro and 18 Pro Max. The camera sensor signs the raw pixels the instant it captures them, before any software can touch them. Apple's servers then turn the raw file into a JPEG inside Private Cloud Compute and sign it again with a post-quantum signature. Anyone with iOS 27, iPadOS 27 or macOS 27 can check the signature and see if the photo has been revoked. This matters because AI image tools now make fakes that look real, so a photo that looks real proves nothing on its own. A reference image proves one narrow thing: these exact pixels came out of a real iPhone sensor at about this time. It does not prove the scene was true. Capture is off in the EU and China at launch. If you handle photos as evidence, start asking for the reference image, not a screenshot.

Level UpIf you or anyone on your team has an iPhone 18 Pro, turn the mode on this week and take one test photo. Settings, then Camera, then Reference Image, then Add Reference Mode. In the Camera app, swipe to Reference and shoot. The reference version shows up next to the normal photo in Photos, and iOS 27 checks its signature every time it is shown. Then decide where a signed photo would save you an argument: damage claims, site visits, hardware serial numbers, before and after shots on a ticket. Write that down as a one-line rule for the team. Two limits to know before you rely on it: capture is off in the EU and China at launch, and a reference image proves what the sensor saw, not that the scene was honest. iGeeksBlog: how to turn on Reference Image mode on the iPhone 18 Pro