Latest Insights
Thoughts on prompt engineering, AI development, and the future of coding.
Andrew Ng Calls the AI Doom Wave a PR Campaign. Anthropic Answered With a Speedometer.
Two documents landed the same week: Ng's skeptic position paper and Anthropic's three public pace metrics. Here's how to read the AI slowdown debate without picking a camp.
Muhammad Hanifa Khairurahman · 
Agencies Are Selling Systems Now, Not Hours. That's the Whole Story.
Eight agencies rebuilt around AI show the same shift: sell the system, keep the saved hours as margin. The Seer lab-to-product pattern is the one to copy.
Steal Meta's Agent Security Checklist: How Muse Assumes It Will Get Hacked
Meta's Muse agent assumes prompt injection will succeed and designs the whole OS around it. Six security patterns every agent builder should copy.
Your AI Spend Needs a Receipt: Build an ROI Framework Before the Vendor Builds It for You
OpenAI's new Admin Console analytics can tell finance what your AI spend bought — but a vendor's dashboard is not evidence. Here's the one-task baseline method to build your own ROI framework before renewal season.
The Voice Agent Race: Gemini 3.8 Live vs GPT-Live-1
Google answered OpenAI's GPT-Live-1 in five days with Gemini 3.8 Live taking the top speech-to-speech spot. Here's how to pick a voice model for production agents without getting burned by vendor benchmarks.
Your Prompts Are Code: The Four-Rung Ladder from Doing Your Job to Building the Tool
Andy Crestodina argues the best person to build your teams AI assistant is the expert who already does the job. Here is the four-stage ladder from manual work to shared workflows, and how to climb it this week.
SWE-2: Frontier Coding for a Third of the Price, Built on a Free Model
Cognition's SWE-2 lands one point behind Claude Fable 5.1 at 64% lower cost, post-trained from the free Kimi K3 — and DeepSeek's un-retired V4 Pro confirms pricing power at the frontier is eroding fast.
An Hour of Audio Now Costs a Dime. Here's What to Build on That.
Three speech-to-text models shipped in nine days and drove transcript prices to $0.10-$0.54 per audio hour. The workflow that matters now: turning one recording into five content assets for under the cost of a freelance blog post.
ChatGPT Ads Are Buyable Now: A First Playbook for the Answer Box
Amazon DSP now sells labeled ad units beneath ChatGPT answers in a US pilot. What shipped, why search instincts misfire, and how to test the surface without wasting budget.
Apple Can Now Prove Your Photos Are Real. Can Your Content Pipeline Survive That?
iPhone 18 Pro's Reference Image moves photo verification into hardware. Here's what provenance checking does to UGC campaigns, influencer proof shots, and your content pipeline.
Uber Just Showed Every Profession How to Fight the Automation Timeline
The first big case study of an incumbent using regulation to slow AI displacement from the inside, and what knowledge workers should copy from it.
The Best Medical AI Doesn't Replace Doctors — It Reads the Tests They Already Ordered
Imperial College's ECG model flags hidden heart disease in tests clinicians already order, for near-zero marginal cost. Here's how to read the honest numbers behind the headlines — and why triage, not diagnosis, is medical AI's real wedge.
One Video Is the New Robot Training Manual
Skild's S1 robot model learns unseen 10-minute tasks from a single video with no fine-tuning — and drops the cost of onboarding a robot task toward the cost of a phone clip.
Your CMO Thinks AI Is Working. Your Team Disagrees.
Jasper surveyed 1,400 marketers: 91% now use AI, but ROI confidence fell to 41% and the CMO-vs-team confidence gap is 61% vs 12%. The problem in 2026 isn't adoption, it's governance and measurement.
Stop Prompt-Praying: Atlas Puts You in the Director's Chair
World Labs' Atlas replaces prompt roulette with actual camera control: one model generating video, 3D scenes, and bullet-time from a handful of phone photos. Here's what that changes for marketing teams.
When the Mockup Becomes the Product: Runway's Solaris and the End of Design-to-Code
Runway's Solaris renders interfaces as live video with no code behind them. What that means for the design-to-code budget, CRO, and web analytics.
Analyze 90-Minute Videos for a Third of the Cost: Gemini's Agentic Video Mode
Google's agentic video understanding cuts token use by up to 88% and cost by up to 66% on long-form video analysis. Here is what that changes for content teams, and how to switch it on today.
750 Tokens a Second: Speed Is Deciding Which AI Agents Survive
OpenAI and Cerebras previewed an 11x-faster API tier this week, and Google and Nvidia shipped speed-focused releases the same week. Why latency now decides which AI agents are usable in production, and how to evaluate it before your next tool purchase.
AI agents now hold their own wallets. The 3D body is the least interesting part
An open-source stack released Aug 29 gives AI agents a 3D body, persistent memory, and their own on-chain wallets that pay other agents in USDC per call. Here's what's real, what's checkable, and why the wallet matters more than the avatar.
Claudeforce and the Default Model Wars
Salesforce and Anthropic shipped a 37-skill sales plugin, but the real news is Claude becoming the default model in Slack. Distribution, not benchmarks, decides who wins.
The Cost of Listening Just Collapsed
IBM's Granite Speech 5.0 Turbo CTC transcribes 3.5 hours of audio per second of GPU compute. Here's what near-free transcription does to voice agents, podcast search, and meeting notes.
How to Stop Renting AI: The Thomson Reuters $40M Playbook
Thomson Reuters spent $40M building its own legal AI on Qwen — and the latest training run cost just $450k. The build-vs-rent math for content businesses has flipped.
Better Prompts Won't Fix Your AI Slop. Governance Will.
Forrester says 68% of buyers distrust AI-made content. Jasper's new governance framework explains why better prompting can't fix that — and what to build instead.
Higgsfield Faceless Studio: The Channel Factory Has a Memory Now
Higgsfield's Faceless Studio turns a channel description into finished five-minute faceless YouTube episodes in about 30 minutes. The real feature is channel memory. Here's the credit math and the monetization fine print.