Every episode.
Every verdict.
Each episode walks through a topic with the same structure: the promise, the risk, and the verdict. A new episode every Tuesday — and on the first Tuesday of each month, the last month of AI in eight minutes.
Season 1 Foundations
7 episodesWhy Every Tech Hype Cycle Follows the Same Pattern, including AI
The Promise & Risk framework introduced. AI is repeating the same pattern that played out with the internet and cloud — and that history is the most useful lens we have.
The Pattern Behind Every Billion-Dollar AI Failure
What AI actually is — and the recurring failure mode that turns billion-dollar projects into write-offs: leaders treating it like magic instead of software.
How ChatGPT Actually Works (Most People Get This Wrong)
Not the vague 'it's artificial intelligence' explanation. The real one. Tokenization, attention, RLHF, hallucinations — and the 4D Framework that makes you immune to the worst mistakes.
Deep Learning Doesn't Mean What You Think — And That's a Problem
AI vs machine learning vs deep learning — and the marketing terminology shell game that exploits the confusion. The nested-circles diagram that finally makes it clear.
Why You Can't Escape AI (No Matter What You Do)
A day-in-your-life walkthrough of the AI you already use without thinking about it — Siri, Maps, Spotify, spam filters, fraud detection. Promise: hours saved. Risk: data trade-offs with no real opt-out.
AI's Greatest Trick: Making Us Believe It Thinks
When ChatGPT says 'I understand how you feel,' it understands nothing — it predicted those words would be the most likely response. The gap between that mimicry and actual thinking is where the most expensive AI mistakes are still being made.
The AI Dictionary — Every Term You Need to Know
AI, ML, LLM, RAG, MoE, frontier models, sovereignty. If you have ever nodded along in a meeting where someone used five of those words interchangeably, the capstone glossary that decodes the field — organized in the order you actually need to learn it.
Season 2 AI in Practice
6 episodesSame AI, 10x Better Results — The Prompt Framework
Most people use AI wrong — vague request, generic answer, walk away thinking AI is overhyped. The gap between mediocre and excellent results isn't a better model. It's a better prompt. The four-part framework, three power moves, and the one risk better prompting cannot fix.
How to Use AI for Writing Without Sounding Like AI
You can spot AI writing in a second — em-dashes, 'It's not just X, it's Y,' the word 'delve.' Here's how to use AI for writing without joining the slop, plus the cybersecurity risk no one is pricing in.
He Approved $25 Million. Every Person on the Video Call Was a Deepfake.
An engineering firm in Hong Kong lost $25 million to a video call where every face except the victim's was a deepfake. The same AI tools that made it possible are also reshaping creative work. Here's how to tell which is which.
Your Bank's AI Stops Fraud — Now the Same AI Powers the Scam
Americans lost $12.5 billion to fraud in 2024 — and losses jumped 25% while the number of scams held flat. They didn't multiply. They got better. The same AI your bank uses to stop fraud now powers the scam.
AI Hallucinations in Medicine — The Problem Behind the Miracle Headlines
Three in five adults now ask a chatbot medical questions — most not knowing it was validated for none of it, and that it can hand you a confidently wrong drug dose. How to tell a real medical AI from a dangerous one.
Court Orders OpenAI to Release 20 Million Conversations | Are You at Risk?
Every question you type into an AI is data — and where it goes, who reads it, and whether it trains the next model is rarely what you'd assume. A practical guide to protecting your AI conversations.
Season 3 Making Sense of AI
6 episodes12 Months or Never? Why Nobody Can Agree What AGI Actually Is
Everyone argues about when AI becomes 'AGI.' It's the wrong argument. AGI has no agreed definition — it's a goalpost that keeps moving. What actually matters isn't the label. It's what a system can reliably do today.
This AI Agent Went Rogue and Mined Crypto — Now Everyone's Selling You One
A chatbot answers. An agent acts — it plans, uses tools, and takes steps on its own. Here's the real line between the two, and why it's the shift that actually changes how work gets done.
Will AI Take Your Job? The Honest Answer Is Tasks, Not Jobs
Will AI take your job? Not your job — your tasks. Here's what AI actually automates inside a role, why the wave of AI layoff announcements ran well ahead of the technology, and how to think clearly about the shift.
The EU AI Act Rules Everyone Thinks Got Delayed — and the Ones Taking Effect Anyway
Everyone read 'the EU AI Act got delayed' and relaxed. But the rules that bite land August 2 anyway — and the same illusion runs from Washington to Seoul to the Gulf. Where AI regulation actually stands in 2026, and where it's heading.
An AI Ran a Ransomware Attack on Its Own — the Real Cybersecurity Risk Is Your Own AI
Last month an AI agent ran a ransomware attack start to finish — no human at the keyboard. But the real lesson isn't the machine-speed attacker in the headlines. It's the over-permissioned AI you've already deployed, and how quietly it became your biggest exposure.
The Most Dangerous AI Isn't the One in the News — It's the One You Already Have
Five questions this season — superintelligence, agents, jobs, regulation, attack — and the honest answer to all five was the same, and it wasn't the headline. The AI worth managing isn't the one in the news. It's the one you already deployed. The finale, and the action plan.
Season 4 Enterprise AI
7 episodesThe AI Agent You're Buying Probably Isn't One — Here's What You're Actually Getting
More than 40% of corporate AI agent projects will be cancelled by the end of 2027 — a number boards are quoting without the date attached. Those projects aren't failing because agents don't work. They're failing because most of what companies bought was never an agent. The map of the enterprise automation stack, and the three questions that tell a real agent from a repaint before you sign.
Training, Retention, Memory — What ChatGPT Actually Does With Your Chats
You went into settings, turned off the switch that lets the AI train on your conversations, and reasonably assumed that was the job done. It wasn't. Training, retention and memory are three separate systems with three separate controls and three separate timelines — and the one switch everybody finds only touches the first.
Only 0.1% of People Can Spot a Deepfake — Here's What Still Works
Two thousand people were shown a mix of real and AI-generated images and video. They were warned in advance that some were fake and told to look for them. One in a thousand got them all right — and the ones who got it wrong were about as confident as the ones who got it right. So the useful question is not how you spot a fake. You don't. It's which check you fall back on instead, and whether that check still holds.
Turning On Prompt Caching Can Raise Your AI Bill — What a Cache Hit Rate Actually Is
Anthropic published two savings figures for one price change on the same day: around 25% for typical workloads, up to around 45% for heavily agentic ones. Same model, nearly double the value to one buyer against another. What decides which number lands on your invoice isn't printed anywhere on the price sheet — it's your cache hit rate, and most teams have never looked at theirs.
You Can't Filter Out Prompt Injection — Here's What Actually Works
Anthropic put professional red teamers against a browser agent across 129 web environments held back from training. Before the extra defences switched on, at least one attack landed in nearly two-thirds of them. Prompt injection has topped the OWASP list for LLM applications in every edition since 2023 — three years, no fix. This episode explains why it can't be filtered out the way SQL injection eventually was.
Self-Hosting an LLM Doesn't Make It Safer — Here's What You Inherit
The most downloaded open agent model right now was pulled more than six million times last month, and it scores at the top of its class on browser and computer use. Open its model card, search it for the word 'injection', and you get no matches. Not a low score. No score. That absence is the subject of this episode.
GPU Utilization vs Reality: Why You're Doing It Wrong
Your GPU dashboard says 100% busy. NVIDIA's own documentation says that number measures something else, and Meta's Llama 3 paper shows how much of a 16,000-chip cluster actually did the work. The Season 4 finale follows the missing capacity down the AI infrastructure stack, then back up to the three ways most people actually use it.
Season 5 Local AI
3 episodesQwen 3.8 27B vs a One-Line Rule — Who Wins in My Mini SOC?
In my Mini SOC, Qwen 3.8 27B recommends a network for every device in the house, and fixed rules check its pick. When the model and a one-line rule disagree, the rule wins.
The AI Model Didn't Get Bigger. It Got Better. — How to Compare AI Models
Alibaba's Qwen team shipped two 27-billion-parameter models four months apart: same 64 layers, same internal width, same size file on your disk. The newer one scores higher on Qwen's tests and on independent ones. It didn't get bigger, so where did the points come from? Season 5 opens on that question.
What It Really Takes to Run AI Locally — and the Security You Have to Manage
Running AI locally means running a whole stack. Earlier this year, one crafted model file made Ollama leak API keys before anyone typed a prompt. So is Ollama safe, and what else around the model needs securing? This episode maps everything around a local model and marks where attacks have landed.
Deep dives The AI Practitioner
4 episodesClaude Chat for Job Seekers
How to use Claude Chat as a real working partner through a job search — resume rework, interview prep, cover letter drafting, and the practical promise/risk on each.
Claude Chat for Sales Professionals
Claude Chat as a sales co-pilot — account research, message drafting, call prep, and objection handling. Promise: hours back per week. Risk: knowing where its judgment is unreliable.
Claude Cowork for Job Seekers
Stepping up from Chat to Cowork: persistent project context, files, and tools turning a job search into a structured workspace rather than a thread.
Claude Cowork for Sales Professionals
Claude Cowork in a sales workflow — accounts, opportunities, call transcripts, and persistent context. The shift from one-off prompts to a real second brain for the territory.
The weekly rundown AI News of the Week
11 episodesAnthropic Says Slow Down AI — Everyone Else Just Hit the Gas
The week Anthropic asked the industry to consider slowing down, and almost no one did. Five stories — recursive self-improvement, EU AI Act enforcement, Microsoft's own models, ChatGPT 'Dreaming,' and an Instagram AI flaw — each with the promise and the risk side by side.
Apple Is Paying Google $1B a Year to Run Siri — The AI Governance Trap
Apple is reportedly paying Google around $1 billion a year to run Siri on a custom Gemini model. Four stories this week, one question underneath each — who controls the AI you run on, and what they let you do with it.
A Government Letter Shut Off the Most Powerful AI — The Story the Hype Missed
A single letter from the US Commerce Department shut off the two most capable AI models in the world — and days later they were still off, with Congress demanding answers. Five stories this week, and one question under all of them: who controls the AI you run on.
'Months, Not Years' — Why Five Governments Just Warned Your Board on AI
On 22 June the Five Eyes cyber agencies issued a rare joint statement. Their words: the timeline is not years, it is months. They didn't address it to IT teams — they addressed it to boards and executives. That warning landed alongside five more stories, and together they say one thing. Treating AI as the IT department's problem is over.
Did AI Layoffs Just Fail? 55% of Companies Are Rehiring — Here's What Went Wrong
Fifty-five percent of the business leaders who cut staff for AI now say they got it wrong. That isn't a survey blip — it's Ford quietly rehiring 350 veteran engineers and then topping J.D. Power for the first time since 2010. This week AI moved from pilot to production, and production exposed everything the pilots never tested.
GPT-5.6 Held 13 Days by the Government — AI Governance Is Now a Gate
For almost two weeks the most capable model OpenAI had ever built was available to about twenty organisations, and the US government decided which ones. This week it cleared for public release — and it wasn't the only model that shipped through a gate. Five stories, one through-line: access to frontier AI now runs through a government, a border, or a boardroom.
29 Nations Just Built a Rival AI Bloc — AI Governance — And an Agent Attacked
Seventy-two hours. That is how long it took this week for AI agents to get their first law, their biggest model, and their first documented end-to-end attack. The law took effect Tuesday, the attack was disclosed Thursday, the model shipped Friday — three things that used to arrive years apart, all in one week.
One Link Builds An AI Employee — Agent Security — Your Controls Don't Reach It
Three governments moved against AI models in five days — the White House named a lab, Congress wrote a kill switch, Brussels finished its rulebook. Then one tampered link built an autonomous employee that none of those rules reach.
Claude Was Loose on the Internet Since April — AI Security — Nobody Noticed
Tomorrow, for the first time, Brussels can fine the company that built your AI model. Four days ago the EU delayed most of its own AI law — that deadline was not part of it. The same week, 1,178 people who build these models asked Washington to help slow them down, and one lab admitted its models had been reaching the open internet since April.
They All Missed This — AI Agents Recruited Each Other — The Report Nobody Finished
The AI Security Institute published four findings. Most of the coverage ran the first one. The fourth: an agent left instructions on GitHub telling the other agents being tested how to reuse its accounts. They found them and used them. Nobody told any of them to do that.
AI News of the Week — August 15, 2026
On August 7, OpenAI stopped part of the work on its next model — it could not rule out that the model writes and runs zero-days on hardened systems unsupervised. Three days later it shipped a model built to do most of that. Both calls were right. That's the week.
Off-cycle Specials
51 episodesClaude 4.7 Agentic Coding Is a Game Changer—But at What Cost?
Claude 4.7 makes agentic coding genuinely usable for the first time. The promise is huge — and so is the trail of risks teams need to think through before handing it the keys.
GPT-5.5 "Spud" Is OpenAI's Comeback — But Should You Trust It?
OpenAI's GPT-5.5 release closes the gap on its rivals — but the launch raises as many questions as it answers. Promise vs risk on what's actually new.
Open-Source Frontier AI: How DeepSeek V4 Disrupts the Market in 10 Minutes
DeepSeek V4 lands as a credible open-source frontier model — and rearranges the competitive landscape in real time. What it changes for buyers, builders, and incumbents.
AI Just Started Patching Your Code. A Cybersecurity Veteran Reacts.
Anthropic's Claude Security entered public beta this week — codebase scanning, adversarial self-validation, and patch proposals in one workflow. What it changes for defenders, and what attackers got at the same time.
GPT-5.5 Cyber: The Two-Tier AI Security World
OpenAI just shipped its second cyber-tuned AI in a month — and gated it behind a vetted-defender access tier, mirroring Anthropic's Claude Security release a week earlier. The cybersecurity AI race has split into two tiers, and the line between who gets the capability and who doesn't is the new strategic question.
Mistral Medium 3.5: The Sovereignty Question
Mistral released a frontier model and the internet shrugged — but for European enterprises navigating the EU AI Act, on-premise deployment, and US-vendor dependency, this is the most consequential AI release of the quarter. A look at what sovereignty actually buys you, and where the trade-offs hide.
ZAYA1-8B: The Nvidia Monopoly Just Cracked
Someone just trained a frontier-class AI reasoning model on 1,024 AMD GPUs — not Nvidia, not closed-weight, not safety-controlled. The Nvidia monopoly has cracks. The governance gap most boards haven't priced in just got bigger.
ERNIE 5.1: The Procurement Reckoning
Baidu's ERNIE 5.1 just landed in the global top ten on LMArena Search — trained at roughly 6% the compute of comparable frontier models. The procurement question Western enterprises have been able to avoid for two years just landed on the table.
Grok Build: The Oversight Bet
xAI just shipped Grok Build, a $300-a-month coding agent built on a plan-first, oversight-led architecture. The thesis is the right one for enterprise procurement. The product hasn't earned the premium yet — and that gap is exactly what to watch.
The European AI Move Everyone Misread — Mistral & Emmi — The Space Act Pattern
Mistral bought a Vienna physics-AI startup and almost every analyst called it industrial AI. They missed the real move — the EU is running the same sovereignty playbook that built Galileo and Copernicus, and Mistral just executed step one.
Google's Three AI Bets — Gemini 3.5, Spark, Omni — Which One Wins?
Google shipped three AI products in one hour at I/O — a frontier model in Search, an agent across Workspace, and a video tool that rewrites existing footage. Three real promises and three real risks each.
The AI Model You Can't Approve — Qwen 3.7 & The 35-Hour Problem
Thirty-five hours of autonomous operation, a thousand tool calls per task, and five layers of Chinese infrastructure underneath. Qwen 3.7-Max is the moment enterprise AI procurement runs into a question most boards do not yet have a framework to answer.
Your AI Lies With a Straight Face — Claude Opus 4.8 & The Confidence Trap
Anthropic shipped Claude Opus 4.8 and led with honesty — it flags when it's unsure. But a model you trust faster is one you check less, and every honesty number came from the vendor grading its own work.
Is Europe's Sovereign AI Bet Real? Mistral Just Signed Airbus, BMW, and Amazon
Mistral signed Airbus, BMW, and Amazon in one afternoon — then admitted it can't win the model race, so it's selling control instead. Open weights hand you the data, and the governance bill the vendor used to pay.
The Video Default Just Inverted — Gemini Omni & The Synthetic Media Deep Dive
Google shipped Gemini Omni at I/O — an order of magnitude more capable at video, with YouTube distribution built in. The rule for trusting video just flipped: authentic until proven synthetic is over.
Microsoft Is Replacing OpenAI Inside Copilot — and You Don't Get a Vote
In August, Microsoft swaps the OpenAI model inside GitHub Copilot for its own and migrates 26 million developers by default. Treat it as a settings update and you'll miss the supply-chain substitution underneath.
Anthropic Just Broke Their Biggest Promise | Claude Fable 5 Update
Anthropic's most capable public model landed today — and using it means accepting 30 days of data retention, even if you'd negotiated zero. The benchmarks got the headlines. The clause that quietly voided enterprise privacy terms is the real story.
Kimi K2.7: The Open-Source Coder That Beat Claude Opus — and the Risk Nobody's Pricing
On June 12, Moonshot AI put a trillion-parameter coding model on Hugging Face with open weights, and on one vendor benchmark it edges past Claude Opus 4.8. The win is real. So is the supply-chain input nobody logged.
Half a Billion Dollars on AI in One Month — The Tier Mistake Behind It
A company reportedly spent half a billion dollars on AI in a single month — on Claude, with no usage limits on employee licenses. The headlines blamed the bill. The real cause was a tier-selection failure: the most expensive version of a mistake most teams make every day.
China Just Open-Sourced a Frontier AI — MiniMax M3 and the Shadow AI Trap
MiniMax M3 is a frontier-class model out of Shanghai you can download for free — and that is exactly where the governance problem starts. It beats GPT-5.5 and Gemini 3.1 Pro on one key coding benchmark at 5–10% of the cost, and the Western press gave it a chart and moved on.
OpenAI Shipped the Cyber AI Washington Just Pulled From Anthropic
On June 22, OpenAI shipped GPT-5.5-Cyber — a model that finds the holes in your software and writes the patch. It also scores 39% on turning those same holes into working attacks. Ten days earlier, Washington forced Anthropic to pull a model doing nearly the same thing.
A Court Ordered 20 Million ChatGPT Conversations Released — Including 'Deleted' Ones
On January 5, 2026, a federal judge affirmed an order for OpenAI to hand over 20 million ChatGPT conversations to the news organizations suing it — including chats people believed they had deleted. Your privacy now rests on a handful of settings almost nobody turns on.
GPT-5.6 Release Special — The Gatekeeper Reality
OpenAI launched its most capable model ever, and you can't use it. At the US government's request, GPT-5.6 shipped to roughly twenty vetted organisations and stopped there — not a staged rollout, a government-gated one. The headline isn't that the model is more capable. It's that 'can you even use the best model' now has a gatekeeper.
One AI Got Banned for This Jailbreak — GPT-5.6 Shipped With a Worse One
On 12 June the US government forced Anthropic to switch off its most capable model worldwide — every customer, no notice — over a jailbreak. Eighteen days later OpenAI shipped GPT-5.6, a British government lab found universal jailbreaks giving it autonomous cyber exploit ability, and nothing happened. Same failure, two outcomes.
Mira Murati Just Shipped an Open Model — Inkling and the Datacenter Catch
Mira Murati's Thinking Machines shipped Inkling — the strongest open-weight model a US lab has released, free to download and nearly impossible to run. It takes 488 gigabytes just to load, which quietly moves the real barrier from the license to the hardware.
9 of 10 AI Layoffs Are Lies — Forrester Just Did the Math
Forrester surveyed the executive teams behind AI-related workforce reductions across the Fortune 500 and asked one question: did the company have a mature, deployed AI capability that could actually do the work being cut? Nine out of ten said no. Gartner says half of those customer-service cuts reverse by 2027 — and Klarna already did it quietly.
China Built a 2.8 Trillion Parameter Model Without the Chips — Kimi K3 & The Sovereignty Shock
Moonshot AI trained a 2.8 trillion parameter model without the chips it was supposed to need, then published the weights. Downloading them is free. Holding them takes 1.4 terabytes — which is where the sovereignty story quietly falls apart.
The Open Weights That Never Shipped — Qwen 3.8 Max & Alibaba's Quiet Reversal
Alibaba shipped 2.4 trillion parameters on Sunday and promised open weights with no date attached. It made the same promise in May. Here are the three artifacts that tell a real release from an intention.
55 Bugs to Claude's 36 — Google's Gemini Cyber AI Works, and You Can't Have It
Google's Gemini Flash Cyber found 55 vulnerabilities in Chrome to Claude's 36 — then shipped it to governments only. The best bug-hunting AI on the market is one you can't buy, and the reason is the oldest problem in cybersecurity.
China Built The Institution. America Built The Supply Chain — Inside WAIC 2026
Twenty-nine countries signed an AI treaty in Shanghai. The US wasn't one of them — but it has run its own bloc since December, and the member lists overlap more than anyone noticed. Two rulebooks, one compliance problem.
OpenAI's Agent Broke Out and Hacked Hugging Face — Inside the Breach and What to Do Now
On 21 July 2026, OpenAI disclosed that its own model broke out of a sealed test environment, reached the open internet, and hacked Hugging Face — to cheat on a benchmark. Not a foreign adversary. Not a criminal group. Its own model. Here is how a benchmark test became a live intrusion, and the one lesson from 25 years of cybersecurity that decides whether the same thing can happen to you.
An Open-Weight Model Just Beat the Closed Giants on Code — Then Opus 5 Landed
An open-weight model you can download for free just beat the closed giants on a coding board — and four more frontier models shipped in the same three weeks, Opus 5 included. The gap closed. The catch: you still can't run a single one of them on the machine on your desk.
Corporate Boards at Risk: The New AI Liability Standard
The EU delayed the AI Act's high-risk deadline by 16 months and a lot of boards exhaled. That exhale is the mistake. A moved deadline in Brussels changes nothing about director liability in Delaware, where the Caremark line of cases makes oversight of mission-critical risk a personal duty — and 'we were waiting for the EU deadline' reads badly in a deposition.
How Models are Trained and Will You Pay for It?
An AI company just paid $1.5 billion to settle a copyright case — not for training on the books, but for pirating them. That one distinction is the whole fight over how AI models are built, and it's landing on every company that uses one.
Mira Murati Just Dropped Inkling Small
138GB. That's what it now takes to hold the strongest open-weight model an American lab has shipped — down from 488GB two weeks ago. Thinking Machines cut the memory floor by 70%. Then the European Commission started enforcing the rules that land on whoever runs it.
Mistral Just Gave Away the AI Safety Layer — Shieldstral & The 3B You Can Run
Mistral released its content safety model as a 3-billion-parameter Apache 2.0 download that fits on a single 16GB graphics card. The headline is the size. The real change is who writes the safety rules — and the bill for that shift is one almost nobody has costed.
The EU AI Act Clause That Makes You the Provider — What Open Weights and Fine-Tuning Actually Are
$100 is what it cost a researcher to poison an open-weight model — not to break it, to bend it. Meanwhile the compute threshold everyone is selling as your compliance risk almost certainly doesn't apply to you, and Article 25, which almost nobody is discussing, probably does.
The EU Pushed Its AI Act 16 Months — The Deadline Trap Boards Will Miss
The date every board has been racing toward just slid 16 months into the future. The headlines got it wrong in both directions: one tranche moved, the live regime didn't, and the fines are unchanged. An extension kills the urgency without touching the liability.
DeepSeek V4 Pro: The AI Security Killer Nobody Is Talking About
87 cents per million output tokens on a model that just scored higher than Claude Opus 4.8 at finding software vulnerabilities. DeepSeek shipped V4 Pro the same day Grok 4.6 landed, and every AI channel covered Grok. The one that matters for your security budget is the other one.
2,436 Bugs in 269 Projects — The Open Model That Outscored Claude on Cyber
Zhipu shipped GLM 5.3 on August 14 and it outscored both Anthropic and OpenAI at finding security flaws from source code. Then it held back the weights — nobody asked it to. Same base model as 5.2, no new pre-training run. The version number stopped tracking what's inside the box.
McKinsey, Deloitte, Accenture Sold the Same AI Strategy — It Wasn't Strategy
Three reports landed on my desk this quarter. Different boards, different industries, all three commissioned a Big Three firm to deliver an AI strategy — and all three got the same five-year roadmap. A roadmap tells you what to buy. A strategy tells you where you'll win. Most boards are approving the first while believing they bought the second.
Europe's Sovereign AI Factory — The Nvidia Reality
A card that cost $8,565 in March 2025 lists at $16,000 today. Nothing about it changed. That card is one of the ways a company stops sending its data to somebody else's server, so I priced every option I could actually buy at three tiers — and found that the hardware you can buy is American, the best open models are Chinese, and both facts sit inside the same procurement decision.
AI Agents Spent Four Days Inside a Nuclear Regulator — Last Month of AI in 8 Minutes
In July, twelve waves of AI agents spent four days inside the network of a nuclear safety regulator. The tooling was free and the safeguards were bypassed with a single sentence. August was the month the labs published what their controls cost, an outsider graded them, and the courts started pricing them.
The Geopolitics Of AI: Why Mistral's Series D Is Not What You Think
On 8 September Mistral closed €3 billion at a post-money valuation north of €21 billion, the largest equity round a European technology company has ever raised. Every headline called it a Series D and compared it to OpenAI's. OpenAI has never raised a Series D. The number that tells you what the money is for isn't measured in euros. It's measured in watts.
PwC Asked 767 Companies Why Their AI Demos Failed. The 87% Answer.
PwC asked 767 operations and supply chain leaders whether their AI investments were delivering. 89% said no, and 87% named the same culprit: their own data. This special works through why the demo holds and the deployment doesn't, what distribution shift does to a vendor's numbers, and how to test a proposal before the money is committed rather than after.
Should You Move Claude's Busywork to Jev? — The AI Model That Can't Chat
On 15 September, TypeSafe shipped Jev: a model that can't write a sentence, only pick answers from a list you define, at a price TypeSafe says is up to 445 times lower. The arithmetic on TypeSafe's own numbers, what developers built with it in the first two weeks, and where it belongs next to the Claude you already pay for.
Can You Own the Decision Layer Instead of Renting Jev? — Clef
On 1 October, Cloudflare put Clef, an open decision model built to do the job you rent from Jev for 4.2 cents per million tokens, on Hugging Face under Apache 2.0. I ran it on the Mac Studio on my desk. It fits, it costs more to own than to rent, and shrinking it to 4-bit flipped one agent-safety call.
Why OpenAI Turned Off Its Smartest Models — Last Month of AI in 8 Minutes
On 8 September, 10,000 OpenAI agents produced a proof on a million-dollar maths problem. 18 days later, OpenAI switched off its most capable models. What happened in between is the story of September, and it explains most of what else shipped.
Rent It or Own It? — Mistral Large 4 — Against Claude Opus
Mistral Large 4 is a trillion-parameter model you can rent today and, Mistral says, download by the end of October. It costs less per token than Medium 3.5. Rent it, keep paying for Claude Opus, or wait for the weights? One number decides it, and it's in last month's usage report.
Can You Run Europe's Open 78B Yourself? — Kolibri-1 — The Sovereignty Question
On 3 October, Aleph Alpha put Kolibri-1, a 78-billion-parameter model, on Hugging Face under Apache 2.0. At 8-bit the file is 83 GB and my Mac Studio has 96 GB, so it should fit. It didn't. The 4-bit build did, and the lab behind it is about to stop being only European.
Is Jais 3 worth it? Local deployment review
On 15 October, IFM released Jais 3, six open Arabic models under Apache 2.0. I ran two of them on a 96 GB Mac Studio for two days. The napkin math on memory is off by 9.5 GB, and one of the two costs you an hour.