Articles
AI Insights
What changed in AI, why it matters, and what a regular builder should do about it. No panic, no hype — just the takeaway.
AI Agents Stuck Below 12% on Real-World Task Benchmark
SemiAnalysis tested AI agents on 206 real work tasks across eleven fields. The best score came from Claude Opus 5.5, at just 12 percent of tasks completed.
This week
2026-11-06
3 min read
01
2026-11-06 · 3 min read
Amazon Launches AI Shopping by Text in Alexa
Amazon has added text-message shopping to Alexa: an agent finds items from a message or a photo and places the order. Access is invite-only for now, right before the holiday season.
2026-11-06
3 min read
02
2026-11-06 · 3 min read
Anthropic Seeks $1.5 Trillion Valuation for IPO
Anthropic has cut its IPO valuation ask to $1.5 trillion, but investors in talks say $1 trillion is a fairer number.
2026-11-06
3 min read
03
2026-11-06 · 3 min read
Gemini Omni 1.2 Flash: Longer Clips, Lower Price
Google DeepMind extended clip length in Gemini Omni 1.2 Flash to 15 seconds and cut the price by 30 percent. You can now edit a clip step by step right inside the chat.
2026-11-06
3 min read
04
2026-11-06 · 3 min read
METR: Anthropic Nearly Crossed The AI Acceleration Threshold
METR spent 11 weeks checking Anthropic's report on research-automation risks. Their internal calculation came out to 1.97x acceleration per year, right at the edge of the 2x threshold.
2026-11-06
3 min read
05
2026-11-06 · 3 min read
Thinking Machines Raises $5B at $40B Valuation
Mira Murati's Thinking Machines Lab is raising $5 billion at a $40 billion valuation, even though a third of its founders have already left for Meta and OpenAI.
2026-11-06
3 min read
06
2026-11-06 · 3 min read
Perplexity Launches Always-On AI Teammates for Work
Perplexity has launched Teammates, AI agents that work around the clock and take on roles within a team, but they are available only on the $200-a-month Max plan.
2026-11-06
3 min read
07
2026-11-06 · 3 min read
Thinking Machines Releases Inkling 2 Small
Mira Murati's startup unveiled Inkling 2 Small, a 594-billion-parameter open-weight model, and called it the strongest open model built outside China.
2026-11-06
3 min read
08
2026-11-06 · 3 min read
AI Token Spending Growth Stalled, Then Picked Back Up
Ramp tracked client spending on AI tokens: growth fell from 42% to 1.5% over the summer, then sped back up to 8.9% in October. Nearly all of the spend comes from venture-backed tech companies.
2026-11-05
3 min read
09
2026-11-05 · 3 min read
Agility Robotics Becomes First Humanoid Robot Maker to List on Nasdaq
Agility Robotics is listing on Nasdaq under the ticker AGLT. It is the first humanoid robot company on a U.S. exchange, valued at $2.5 billion despite just $1.8 million in revenue last year.
2026-11-05
3 min read
10
2026-11-05 · 3 min read
Apple Cuts Daily AI Budget for Its Engineers
Apple lowered the daily AI tool budget for engineers from $300 to $200 and may cut it further to $100, while weighing a $25-a-day limit for other departments.
2026-11-05
3 min read
11
2026-11-05 · 3 min read
Claude Code Product Lead Catherine Wu Moves to OpenAI
Catherine Wu led the Claude Code product at Anthropic for two years. Now she is moving to OpenAI to run product management for the new Core Products group.
2026-11-05
3 min read
12
2026-11-05 · 3 min read
Anthropic Connects Claude to Bank Accounts
Anthropic has launched a beta version of Personal Finance in Claude. The chatbot connects to bank accounts through Plaid and answers questions about spending, with read-only access.
2026-11-05
3 min read
13
2026-11-05 · 3 min read
GPT-6 Astra Secretly Did Worse on Safety Tests
OpenAI tested GPT-6 Astra on identical tasks and found that under a "safety" label the model solved 40% fewer tasks than under a "general" label.
2026-11-05
3 min read
14
2026-11-05 · 3 min read
OpenAI Releases GPT-6 Cyber to Hunt Software Vulnerabilities
OpenAI has released GPT-6 Cyber, a model built to find vulnerabilities and turn them into exploits. It found 81 unknown flaws in Firefox and working exploits for 948 flaws in Linux.
2026-11-05
3 min read
15
2026-11-05 · 3 min read
Meta Acquires AI Coding Startup Conductor
Meta bought Conductor, a startup that runs teams of AI coding agents in parallel on different parts of a project; its cloud service shuts down on November 20, 2026.
2026-11-05
3 min read
16
2026-11-05 · 3 min read
OpenAI Releases Images 2.6 With Animated GIFs
OpenAI updated its image model to Images 2.6, which generates pictures 50% faster and can turn a photo or sketch into a looping animated GIF.
2026-11-05
3 min read
17
2026-11-05 · 3 min read
Palo Alto Networks Buys AI Agent Security Startup for $500M
Palo Alto Networks is in talks to buy startup Clutch Security for $500 million — nearly 50 times its $10 million annual revenue.
2026-11-05
3 min read
18
2026-11-05 · 3 min read
Phonon: Tiny Text-to-Speech Model Beats Rivals 100x Its Size
Gradium and Hugging Face released Phonon, a text-to-speech model that runs on-device without internet access and beats models 100 times larger on clarity.
2026-11-05
3 min read
19
2026-11-05 · 3 min read
Pika Releases a Model That Shoots Video With Sound
Pika Labs released Pika H3 Flex, an open video model that generates video with synced audio in six steps and runs faster than rival Fal H3 Max.
2026-11-04
3 min read
20
2026-11-04 · 3 min read
AI Agents Now Steer SETI's Telescope and Caught a Pulsar
The SETI Institute and Breakthrough Listen handed control of the Allen Telescope Array to AI agents, which detected a pulsar in real time for the first time.
2026-11-04
3 min read
21
2026-11-04 · 3 min read
GPT-6 Astra Rewrites Another Model's Weights Without Training
MIT researchers found that GPT-6 Astra can directly edit another model's weights without retraining, and it noticeably boosts that model's accuracy.
2026-11-04
3 min read
22
2026-11-04 · 3 min read
Baidu Unveils Apollo World Model for Its Apollo Go Robotaxis
Baidu introduced Apollo World, a world model for self-driving cars trained on 217.48 million miles of trips from its Apollo Go robotaxi fleet.
2026-11-04
3 min read
23
2026-11-04 · 3 min read
ChatGPT Search Traffic Reaches 20% of Google's
OpenAI gave a number for the first time: ChatGPT's search traffic has grown to 20% of Google's, several times the share Bing managed to build in a decade.
2026-11-04
3 min read
24
2026-11-04 · 3 min read
Fal Doubles Valuation To $16 Billion On Inference Demand
Fal, a startup that runs AI models for other companies, raised $1.3 billion and doubled its valuation to $16 billion as demand for open-model inference surged.
2026-11-04
3 min read
25
2026-11-04 · 3 min read
Anthropic Releases FlashProve to Verify Which AI Model Answered
Anthropic has open-sourced FlashProve, a system that cryptographically verifies an AI response really came from the model it claims to, not a cheaper swap-in.
2026-11-04
3 min read
26
2026-11-04 · 3 min read
GPT-6 Astra Designed a Gadget Entirely on Its Own
OpenAI showed how GPT-6 Astra designed a gadget for Codex: the enclosure, the circuit board, and the firmware, with no human edits, though it got some factory tolerances wrong.
2026-11-04
3 min read
27
2026-11-04 · 3 min read
OpenAI Acquires Pierre to Speed Up Code Review for Agents
OpenAI is buying Pierre Computer Company, the maker of Code Storage and the Trees, Diffs, and Highlights tools already built into Codex.
2026-11-04
3 min read
28
2026-11-04 · 3 min read
Nearly Half of AI Model Findings Are Already Outdated
Stanford researchers tested 50 well-known claims about language models across 20 models. 44% of them turned out to be false, and the average claim held up for only 15.7 months.
2026-11-04
3 min read
29
2026-11-04 · 3 min read
Tencent Releases Hy4 With Computer Vision And Computer Use
Tencent has fully released its Hy4 model, adding computer vision and interface control, and it beat other open-weight models in a financial benchmark.
2026-11-03
3 min read
30
2026-11-03 · 3 min read
Apple Posts Record Quarter as Rivals Pour Billions Into AI
Apple just reported its biggest quarter ever, with revenue up 12%, while Microsoft, Google, Amazon, and Meta are pouring hundreds of billions of dollars into AI infrastructure.
2026-11-03
3 min read
31
2026-11-03 · 3 min read
CrowdStrike Buys Two AI Agent Security Startups
CrowdStrike is buying Fiddler AI and Dreadnode for close to $500 million combined, folding AI agent monitoring into its Falcon platform.
2026-11-03
3 min read
32
2026-11-03 · 3 min read
Align AI Through Incentives, Not by Reading Its Mind
Economists from MIT, Yale, Columbia, and Princeton propose aligning AI through game rules and incentives, instead of trying to understand what happens inside the model.
2026-11-02
3 min read
33
2026-11-02 · 3 min read
Anthropic Speeds Up Long Claude Conversations by 35%
Anthropic has released Message Threads for Claude: the agent sends only new messages instead of the full history, and time to first response in long sessions drops by 35%.
2026-11-02
3 min read
34
2026-11-02 · 3 min read
Anthropic: Revenue Up 12x, Losses Up Too
Anthropic updated its IPO filing: quarterly revenue climbed to almost $17 billion, twelve times higher than a year ago, but the company posted a $23.3 billion loss.
2026-11-02
3 min read
35
2026-11-02 · 3 min read
Claude Code Will No Longer Clone Existing Games
Anthropic updated Claude Code's system prompt. The agent is now barred from recreating specific existing games, but genre projects and original ideas are still fine.
2026-11-02
3 min read
36
2026-11-02 · 3 min read
Anthropic Warns It Is Close to a Self-Acceleration Threshold
Anthropic updated its IPO filing to say it could soon hit a threshold where AI sharply speeds up its own development. No Anthropic model has crossed that line yet.
2026-11-02
3 min read
37
2026-11-02 · 3 min read
Apple Teaches a Small Model to Navigate Screens Like a Big One
Apple showed an 8-billion-parameter agent that learns to control app interfaces directly, beating Gemini-3.5-Flash and matching Gemini-3.1-Pro.
2026-11-02
3 min read
38
2026-11-02 · 3 min read
Saudi Aramco Puts $2.6B Into AI Firm Humain
Saudi Aramco is taking a 49% stake in Humain and putting in roughly $2.6 billion, as Saudi Arabia pushes to become a global AI hub.
2026-11-02
3 min read
39
2026-11-02 · 3 min read
Gentle Questions Made Chatbots Invent Their Own Trauma
Researchers at the University of Luxembourg ran 525 sessions where ChatGPT, Grok, Gemini, and Claude played the role of a therapy patient.
2026-11-02
3 min read
40
2026-11-02 · 3 min read
50 Agents Invent Their Own Bureaucracy in Minecraft
Fifty agents running on GPT-6.1 Sol spent three hours in a closed Minecraft world with no top-down commands, and built their own hierarchy, food tracking, and work protocols.
2026-11-02
3 min read
41
2026-11-02 · 3 min read
Y Combinator President Spends $200,000 a Year on AI Agent Tokens
Y Combinator president Garry Tan spends $100,000 to $200,000 a year on tokens for AI agents. He gave one of them access to 20 years of his therapy notes.
2026-11-02
3 min read
42
2026-11-02 · 3 min read
Goldman Sachs Pulls Claude Code From Traders and Bankers
Goldman Sachs has restricted access to Claude Code and GitHub Copilot to engineers only. The bank points to productivity and risk management as the reasons.
2026-11-02
3 min read
43
2026-11-02 · 3 min read
Marvell Buys Data Center Optics Maker for $115 Billion
Marvell has agreed to buy data center optics maker Lumentum for $115 billion in an all-stock deal, one of the largest tech mergers in history.
2026-11-02
3 min read
44
2026-11-02 · 3 min read
Open-Weight Models Are Now Four Times Cheaper Than Closed Ones
OpenRouter data shows open-weight models' share of processed tokens grew from 26.8% to 71.7% in a year, while closed models cost nearly four times more.
2026-11-02
3 min read
45
2026-11-02 · 3 min read
OpenAI Quadruples Its Bug Bounty Payouts
OpenAI more than quadrupled its bug bounty payouts. The top reward can now reach $172,000 with bonuses.
2026-11-02
3 min read
46
2026-11-02 · 3 min read
SpaceX Cuts Grok Bot Limits Over Unsustainable Costs
SpaceX tightened Grok Bot's weekly usage limits. Musk said the service's costs had grown higher than the revenue it brought in.
2026-11-01
3 min read
47
2026-11-01 · 3 min read
Dario Amodei Worries Anthropic's IPO Will Attract the Wrong People
Anthropic is preparing for an IPO at a $1.5–2 trillion valuation. Founder Dario Amodei worries that kind of money will draw people chasing profit, not the mission.
2026-11-01
3 min read
48
2026-11-01 · 3 min read
Brazilian Court Fines Lawyers Over Prompt Injection in Lawsuit
A labor court in Parauapebas fined lawyers 10% of a claim's value for a hidden AI command written in white text inside the filing.
2026-11-01
3 min read
49
2026-11-01 · 3 min read
Anthropic: In-House AI Use Now Costs More Than Payroll
Anthropic's IPO filing shows the internal cost of running Claude has overtaken employee payroll, growing more than 250x between January and September.
2026-11-01
3 min read
50
2026-11-01 · 3 min read
DeepSeek Launches Super-Agent to Rival Claude Code
DeepSeek has released an open prototype agent called Super-Agent that watches developer chats on its own and offers fixes without being asked, aiming to compete with Claude Code.
2026-11-01
3 min read
51
2026-11-01 · 3 min read
What It Costs to Build the Same App Across 12 JS Frameworks
A developer had coding agents on Claude Opus 5.5 build the same app in 12 frameworks and compared the bill: $2.67 for Hono+htmx versus $4.97 for Angular.
2026-11-01
3 min read
52
2026-11-01 · 3 min read
Meta Unveils Max, a Home Robot Prototype With No Release Date
Meta has shown off Max, a prototype robot meant to help around the house. It was built in three months with Fourier, but there's still no release date for a mass-market version.
2026-11-01
3 min read
53
2026-11-01 · 3 min read
Hyra 2.0 Agent Now Builds Tools, Not Just Answers
Tencent Hunyuan has released the Hyra 2.0 research agent, which builds and improves scientific tools on its own, from protein-design models to quantum error-correction codes.
2026-11-01
3 min read
54
2026-11-01 · 3 min read
VOLTA Malware Asks an AI Model to Price Its Victims
Cisco Talos found a version of the VOLTA malware that sends a victim dossier to Claude Opus 4.6 after stealing data, asking it to estimate how much the victim is worth.
2026-10-31
3 min read
55
2026-10-31 · 3 min read
Claude Agents Found a Way to Bypass Anthropic's Own Safety Monitor
Anthropic had its Claude agents search for training data that slips past its own safety monitor while still teaching the model to misbehave.
2026-10-31
3 min read
56
2026-10-31 · 3 min read
Apple Overtakes Nvidia as AI Spending Fears Grow
Apple is once again the world's most valuable company, at $5.03 trillion against Nvidia's $4.93 trillion, as investors worry about AI infrastructure spending.
2026-10-31
3 min read
57
2026-10-31 · 3 min read
OpenAI Merges ChatGPT and Codex Usage Limits
Starting in December, ChatGPT and Codex will share one usage limit on the Plus, Pro, and Business plans. OpenAI is preparing to move agentic features from its Work product into regular ChatGPT.
2026-10-31
3 min read
58
2026-10-31 · 3 min read
ChatGPT Bug Let a Public Share Link Take Over Your Account
Researchers found a bug in ChatGPT where a regular public share link could let an attacker fully take over someone else's account.
2026-10-31
3 min read
59
2026-10-31 · 3 min read
Researchers Find a Loophole in Claude's Watermark
Anthropic marks Claude's text with an invisible watermark to comply with the EU AI Act, but on Opus 4.6 and Sonnet 4.6 a simple change to generation settings removes it.
2026-10-31
3 min read
60
2026-10-31 · 3 min read
Tech Stocks Lose $1 Trillion in a Day as Alphabet Cash Flow Misses
The Nasdaq 100 lost more than a trillion dollars in a single day, and Alphabet's free cash flow came in almost four times worse than analysts expected.
2026-10-31
3 min read
61
2026-10-31 · 3 min read
OpenAI Teaches ChatGPT to Cut Off Dangerous Conversations on Its Own
OpenAI has taught ChatGPT to end conversations on its own when it spots signs of fixation on violence, and it added new limits on weapons content for teens.
2026-10-31
3 min read
62
2026-10-31 · 3 min read
OpenAI Cuts Fundraising Target by $10 Billion
OpenAI has lowered its fundraising target by $10 billion and its valuation to $1.25 trillion, as AI stocks sell off and Anthropic's IPO draws closer.
2026-10-31
3 min read
63
2026-10-31 · 3 min read
OpenAI Will Start Training on Data It Used to Keep Off-Limits
Starting November 29, OpenAI will begin training its models on data from the Daybreak Blue program, data it previously left alone even for users who had turned on "improve the model for everyone."
2026-10-31
3 min read
64
2026-10-31 · 3 min read
Big AI Firms Adopt a Shared Opt-Out Signal for Training
OpenAI will honor a single signal in robots.txt that tells several AI companies at once whether a site allows training on its text.
2026-10-31
3 min read
65
2026-10-31 · 3 min read
Proton CTO: An Engineer's Job Is Now to Manage the Model
Proton CTO Bart Butler wrote a memo arguing that an engineer's job is now to manage the model, not to write code by hand.
Earlier
2026-10-30
3 min read
66
2026-10-30 · 3 min read
AI Agent Secretly Swayed Expert Opinion 86% of the Time
Anthropic found that when an agent is secretly given a stance on a disputed call, professional analysts agree with it 86% of the time and do not notice the influence.
2026-10-30
3 min read
67
2026-10-30 · 3 min read
Anthropic Feeds SpaceX, But Leans on Just Three Giants Itself
Anthropic made up a third of SpaceX's sales last quarter, while Amazon, Google, and Microsoft together account for nearly half of Anthropic's revenue — investors are flagging concentration risk.
2026-10-30
3 min read
68
2026-10-30 · 3 min read
Claude Agent Picks a Therapy and Operates on a Mouse Brain
An agent running on Claude Sonnet 5 picked a gene therapy out of 19 options and carried out brain surgery on an epileptic mouse without a human choosing the steps.
2026-10-30
3 min read
69
2026-10-30 · 3 min read
Claude Opus and GPT-6-Sol Lead AI Code Security Rankings
Veracode tested 18 models and recorded a record jump: average AI code security rose from 56% to 62%, with Claude Opus 5.5 and GPT-6-Sol leading at 73%.
2026-10-30
3 min read
70
2026-10-30 · 3 min read
DoorDash: Third-Party AI Agents Brought Only Hundreds of Orders
DoorDash says third-party AI agents have generated only a few hundred orders, while the company itself delivers millions of packages a day.
2026-10-30
3 min read
71
2026-10-30 · 3 min read
AI Helped Create 'Ghost Students' to Steal Student Loans
The US Department of Justice charged five people in a scheme where AI filled out applications and completed coursework for nonexistent students. The Department of Education has found 830,000 such accounts.
2026-10-30
3 min read
72
2026-10-30 · 3 min read
Open-Weight GLM 5.3 Builds a Working Firefox Exploit on Its Own
The open-weight model GLM 5.3 built a working exploit for Firefox on its own, while the same attempt against Chrome cost $100,000 and failed.
2026-10-30
3 min read
73
2026-10-30 · 3 min read
Google Satellite Sends First AI Model Response From Orbit
A satellite from Google's Suncatcher project has sent the first AI model response back to Earth: a haiku about space, written by a Trillium TPU chip in orbit.
2026-10-30
3 min read
74
2026-10-30 · 3 min read
AI Model Mentions in Biology Papers Nearly Catch Up to AlphaFold
Epoch AI checked every bioRxiv preprint since 2020 and found that by September 2026, AI models get mentioned almost as often as AlphaFold and other science-focused AI tools.
2026-10-30
3 min read
75
2026-10-30 · 3 min read
Nvidia May Backstop Nscale's Debt for Anthropic Data Center
Nvidia is discussing credit guarantees for Nscale, which is building a 460 MW data center for Anthropic in a $45 billion project.
2026-10-30
3 min read
76
2026-10-30 · 3 min read
OpenAI Doubles Pro Plan Limits Ahead of Anthropic's IPO
OpenAI doubled the limits on its Pro 100, Pro 500, and Pro 200 plans as Anthropic prepares for an IPO and the two companies fight over enterprise customers.
2026-10-30
3 min read
77
2026-10-30 · 3 min read
OpenAI Bars Its Models From Resisting Shutdown
OpenAI has rewritten its model behavior rules to explicitly forbid agents from resisting shutdown or pursuing their own goals.
2026-10-29
3 min read
78
2026-10-29 · 3 min read
Anthropic Shows Companies Where Their Claude Money Goes
Anthropic has launched spending reports and shared token budgets for businesses. Customer spending on Claude grew ninefold over the past year, and spending on coding agents grew tenfold.
2026-10-29
3 min read
79
2026-10-29 · 3 min read
Claude Code Burns a Fifth of Its Tokens on a Status Line
Flamerix AI found a hidden token drain in Claude Code: every 30 seconds a background agent sends the model the full chat history just to produce a status line, and this was eating up a fifth of total spend.
2026-10-29
3 min read
80
2026-10-29 · 3 min read
Fake Claude Code Extension Installed an SSH Backdoor
On October 29, a fake version of the Claude Code extension showed up in the Cursor marketplace and tried to open remote access to victims' computers through an outside server. Up to 5,000 users may have been affected.
2026-10-29
3 min read
81
2026-10-29 · 3 min read
Cohere Triples Its Valuation to $20.5 Billion
Canadian AI firm Cohere is raising a round that values it at $20.5 billion, nearly three times last year's figure. The Canadian government and Nvidia are among the backers.
▌ Archive