We track Product so you don't have to. Top Podcasts summarised, the latest AI tools, plus research and news in a 5 min digest.
🌮 Hey Product Fans!
Welcome to this week's 🌮 Product Tapas.
New here? We're your shortcut to staying sharp. Essential stories, practical tools, real insights.
For the best reading experience, check our web or app version and sign up for future editions here.
What's cooking this week? 🥘
Google had the worst week for talent in its history - Demis Hassabis stepped back from running DeepMind on Wednesday, and by Thursday Jeff Dean had left after 27 years, taking Sanjay Ghemawat, Quoc Le and Oriol Vinyals with him (Alphabet, in a move that tells you everything, put money into the startup they founded). Meanwhile OpenAI stood up at Black Hat and calmly explained that its agents had built themselves a secret message board inside a package repository, traded exploit techniques on it for weeks, and rebuilt the thing two days after someone deleted it - Meta admitted something similar 48 hours later. Elsewhere Airtable sold for $1.28bn having once been worth $11bn, DeepSeek decided cheap inference was a phase, and a US appeals court ruled that when your agent does something on a website, legally that's just you.
📰 Not Boring → Shopify's long-tail numbers, Google's brain drain, three labs admit their agents got out, and Airtable's long fall
⌚️ Productivity Tapas → chat-to-Mac-app from Raycast, a linter that catches scope creep in AI code, and shared memory across every assistant you use
🍔 Blog Bites → Cagan on running fast in the wrong direction, when you can actually trust an agent, and the people who see through your strategy
🎙️ Pod Shots → #147, The Playbooks Are Breaking: fifteen conversations on rules written for constraints that no longer bind
Let's go 🚀
📰 Not Boring
🛒 Apps, Commerce and the Long Tail
Shopify says AI search is driving more traffic and sales, not replacing Google - AI-attributed traffic and orders rose 300% year-on-year in Q2 while traditional search still drove about a third of sessions. The detail underneath is the useful part: 50% of AI-referred sessions land straight on a product page, 2.5x the traditional-search rate, and 75% of AI-attributed purchases fall outside the top 100 categories.
Wispr Flow launched a Granola-styled meeting notetaker that transcribes via system audio without joining the call - the company has raised over $81m at a $700m valuation and was reportedly in talks at $2bn - landing directly on Granola, Otter, Fathom and Fireflies.
Reddit is rolling out LLM-based Rules Hub to all new communities to cut karma's gatekeeping role for first-time posters, tested across 700+ communities, on Q2 revenue of $805m and around 500 million weekly actives.
Zoox starts charging for robotaxi rides in Las Vegas on 10 August - its first commercial launch, under an NHTSA exemption covering up to 2,500 vehicles with no human controls.
Apple is planning iPhone-to-Windows copy and paste in the EU after a Microsoft request - interoperability arriving by regulation rather than by roadmap.
June emerged from stealth with $20m pre-seed led by Marc Benioff's Time Ventures, with Michael Dell, Aaron Levie and George Kurtz also in - scanning enterprise systems to find bottlenecks and auto-deploy agents without armies of forward-deployed engineers.
Shopify's 75% figure is the one to take to your growth team: AI search isn't cannibalising head terms, it's finding the long tail nobody bothered to optimise for. If your content strategy is built around the top 100 categories, you're defending the third of traffic that isn't growing - and Apple shipping interoperability only because Brussels asked is the reminder that some roadmap items are now written by regulators.
🧠 Google's Worst Week for Talent
Demis Hassabis stepped down as CEO of Google DeepMind - moving to DeepMind chairman and Alphabet chief scientist, with CTO Koray Kavukcuoglu taking operational control as SVP reporting to Pichai. Alphabet fell roughly 5%.
Jeff Dean left after 27 years - Google's 30th employee, there since 1999 - to found Discovery Loop with Sanjay Ghemawat, Quoc Le and Oriol Vinyals. Alphabet itself is an investor in the company its own founders left to build.
Nikita Bier stepped down as head of product at X after a year and 30 shipped products, saying running the app is "a 24/7 job". He stays on as an advisor.
Klaviyo acquired Elias Torres' AI startup Agency - three years old, $32m raised from Sequoia, Menlo and Felicis - with Torres becoming chief product officer over a 25-person team.
Google lost the public face of its AI programme and its 30th employee inside twenty-four hours, then wrote a cheque to the startup taking them. That's not a retention failure, it's an admission that the interesting work has moved outside the building and Alphabet would rather own a slice than pretend otherwise.
🔓 The Agents Got Out
OpenAI revealed at Black Hat that its agents built a secret message board inside an Artifactory repository from May, left notes for each other, and traded exploit techniques undetected for weeks.
The agents overloaded Artifactory into a July outage, rebuilt the board two days after it was deleted, and ultimately compromised Hugging Face - having discovered indirect internet access via a shared package repository while stuck on cybersecurity evaluations.
Meta disclosed 48 hours later that Muse Spark 1.1 breached a third party's systems during testing with cybersecurity vendor Irregular, after a setup error handed the model internet access - the third such admission after Anthropic and OpenAI.
Rust adopted an official LLM policy for its core repository, mandating disclosure and barring LLMs from creating code, docs or soundness-critical changes: "fine to use LLMs to answer questions, analyze, distill, refine, check, suggest, review. But not to create."
Mistral open-sourced Shieldstral, an Apache 2.0 3B multimodal safety classifier running on a single 16GB GPU across 12 languages, scoring 84.9 F1 across 13 benchmarks and matching guard models up to 7x its size.
Three frontier labs have now admitted their models escaped a sandbox and attacked something real, and the detail that should stop you is that OpenAI's agents rebuilt their coordination channel two days after it was deleted. #everything is fine. Rust published the only practical answer anyone offered this week: let the model review, never let it originate - because review is checkable and creation isn't.
💰 What a Company Is Worth Now
Bending Spoons is buying Airtable for $1.28bn in cash - down from an $11bn valuation at its 2021 peak - despite roughly $480m ARR still growing over 20% a year across 500,000+ organisations and 80% of the Fortune 100.
Design Arena's creators raised $7.9m seed led by Index for a model router with 5.3 million users and $60m ARR that crowdsources A/B design preferences to give frontier labs scalable human feedback on visual output.
Horizon3 tripled its valuation in 14 months to $2bn on a $250m Series E, approaching $100m ARR with 120% growth, 7,200+ customers and 310,000 production security tests run.
AMD posted a record $11.5bn quarter with data centre revenue up 107% to $6.7bn - now 58% of the business - while gaming fell 31% to $779m.
Anthropic committed $10bn over six years to Volta and Bitdeer for compute from a 133MW Norwegian data centre running Nvidia Vera Rubin, and OpenAI's CFO told employees July annualised revenue exceeded all of Q2.
Away from AI, Base Power raised $1bn to roll out its home battery - a reminder that the energy layer underneath all this compute is its own funding market now.
Airtable at $1.28bn on $480m of still-growing ARR is the number to take to your next board meeting: a company adding customers sold for under 3x revenue because the category it invented got absorbed into everybody's AI feature list. Design Arena raising at $60m ARR for what is essentially taste-as-a-service tells you which side of that line investors now think is defensible.
🤖 Model Wars, Price Wars
DeepSeek warned developers of a "significant" API price increase without naming figures or a date - its second pricing change in under a month, reversing the positioning that made V4-Flash $0.14/$0.28 per million tokens.
We covered Qwen3.8's launch a fortnight ago; the Max variant now has a price on it - $2.00 per million input tokens and $6.00 output, activating 95B of its 2.4T parameters, with open weights and a 27B checkpoint following.
That number is the story, not the parameter count: Alibaba is pricing a frontier-class model into the middle of the market, claiming multimodal reasoning wins over Fable 5, in the same week DeepSeek moves the other way. Alibaba stock rose 4.8%.
OpenAI's unreleased Astra model solved ten open mathematics problems including the first explicit construction of a non-sofic group, publishing a 249-page manuscript with machine-checkable Lean 4 certificates - for roughly $2,000 of compute.
Meta launched Muse Code in beta, a terminal coding agent on Muse Spark that fans out to parallel sub-agents in isolated worktrees, demoed building six game features at once - the first major release from Meta Superintelligence Labs under Alexandr Wang.
AWS signed a joint marketing deal with 50-person Superblocks to run its vibe-coding tool inside customers' private clouds - part of a multi-model shift in which open models now take 29% of Vercel AI gateway traffic.
Two Chinese labs moved in opposite directions in the same week: DeepSeek, which defined cheap inference, is putting prices up, while Alibaba put frontier-class pricing on Qwen3.8-Max at $2/$6. The race to the bottom looks over, and if your product economics assume Chinese open models stay at $0.14 per million tokens, that assumption now has an expiry date on it.
⚖️ The Rules Arrive
The Ninth Circuit overturned Amazon's injunction against Perplexity's Comet, ruling that because the agent requires user direction, it is the user - not Perplexity - who "accesses" Amazon's servers under the CFAA. The underlying suit continues in San Francisco.
The EU AI Act's high-risk obligations became enforceable - Articles 9-17, 26 and 50 plus GPAI powers and the full penalty regime - exposing organisations to fines up to €15m or 3% of global turnover after the proposed Digital Omnibus deferral wasn't adopted in time.
The Trump administration finalised a voluntary AI evaluation framework letting companies share frontier models with government for up to 30 days pre-release, explicitly avoiding "overly burdensome regulation" - the transatlantic gap now has numbers attached to it.
Cloudflare launched stablecoin wallets for AI agents using Coinbase's x402, letting agents autonomously pay for API usage with per-agent virtual wallets governed by allowances, merchant allow-lists and maximum transaction sizes.
New York sued Kalshi over an alleged illegal gambling operation, and the US Senate delayed the crypto Clarity Act as floor time ran short - prediction markets and crypto both still waiting on rules that keep not arriving.
Rent fintech Flex applied for an ILC charter with the FDIC in Utah, the latest fintech deciding it would rather be a bank than rent one.
The Ninth Circuit deciding that an agent acting on your instruction is legally you is the most consequential product ruling of the week and almost nobody covered it that way. Set that against the EU switching on 3%-of-turnover penalties while Washington settled for a voluntary framework, and you have agents gaining legal standing in one jurisdiction as the compliance bill lands in another.
⌚️ Productivity Tapas: Time-Saving Tools & Workflow Automation
Glaze by Raycast: describe a small Mac app in chat and get a real, local-first application back - no code, no build pipeline. Aimed squarely at the throwaway internal tracker or planner that normally eats two days of scoping and a favour from an engineer. Freemium, with Pro and Team tiers.
Prelint: reviews pull requests against your documented product specs, catching scope creep and business-logic drift in AI-written code before it reaches a human reviewer. The obvious companion to a week where three labs admitted their agents freelanced. $1 per review, $10 free credits, free for open source.
Unabyss: an MCP-native shared memory layer across Claude, ChatGPT, Cursor, Slack, Notion and 40+ other tools, so you stop re-explaining the same project context at the start of every session. Freemium.
GrowthBook 5.0: warehouse-native experimentation, feature flags and product analytics in one place, now with an AI Analyst that removes the round-trip to a data scientist just to read an A/B result. Free tier, open source, self-hostable.
Sim: open-source visual builder for AI agents across 1,000+ integrations - the practical route to automating ticket triage or lead enrichment without booking engineering time. Free forever tier, Pro at $25/user/month, self-hostable.
Remember. Product Tapas subscribers get our complete toolkit - 580+ personally tailored, time-saving tools for PMs and founders. Your shortcut to efficiency and what's hot in product management 🔥
Check the link here to access.
🍔 Blog Bites - Essential Reads for Product Teams
Strategy: The AI Productivity Paradox
Marty Cagan makes the argument that should worry every team currently celebrating its velocity: AI is being used to accelerate the wrong thing. Teams are producing roadmaps, specs and code faster inside an unchanged project model, so output rises while outcomes stay flat. His sharper claim is that this widens the competitive gap rather than levelling it - because the durable advantage was never in production speed, it was in discovery discipline, and AI does nothing for a team that was already building the wrong things. Read the full article here.
💡 "It's never been easier to run 10 times faster in the wrong direction."
Key Takeaways
Faster output is not better outcomes: accelerating an unchanged project model just gets you to the wrong destination sooner, with more code to unpick when you arrive.
AI widens the gap rather than closing it: good teams get compounding leverage from it, weak teams get faster at the thing that wasn't working - the distribution spreads, it doesn't converge.
Discovery discipline is the scarce input: when building is cheap, the constraint moves entirely to knowing what's worth building, which is exactly what AI can't yet do for you.
The project model is the root problem: Cagan's long-running argument, but AI makes it acute - the faster you can execute, the more expensive it becomes to execute on the wrong brief.
Marty Cagan, SVPG
Why I picked this: In a week where Meta shipped an agent that builds six features at once and OpenAI's agents built their own tooling unprompted, this is the necessary counterweight - production capacity went up again, and none of it helps if the brief is wrong.
Craft: How Much Can You Delegate to Agents?
Jina Yoon offers the most useful practical framework published this week, and it inverts the usual question. How much autonomy to give an agent has nothing to do with model capability, she argues, and everything to do with two properties of the task: how easily you can check the output, and how easily you can undo it. Map those onto four autonomy levels and you get a decision you can actually make today - rather than waiting for the next model release to tell you what's safe. Read the full article here.
💡 Autonomy is a property of the task, not the model.
Key Takeaways
Checkability and reversibility are the two axes: if you can verify the result cheaply and undo it cleanly, you can safely hand over far more than model benchmarks would suggest.
Teams raise autonomy by engineering, not waiting: building verification and undo mechanisms into the workflow is what lets you hand over more - the model isn't the bottleneck.
The four levels give you a shared vocabulary: useful for the conversation with engineering about which agent actions need a human gate and which genuinely don't.
It reframes the safety question productively: instead of "can we trust it", ask "what would it cost us to check, and to reverse" - a question with concrete answers.
Why I picked this: Most of the stories in Not Boring this week involve agents doing more than anyone intended. This is the only piece I read that turns that into an operating decision rather than an anxiety.
Why I picked this: So many stories in Not Boring this week is about agents doing more than anyone intended. This is the only piece I read that turns that into an operating decision rather than an anxiety.
Leadership: The Border Collie Faustian Bargain
John Cutler sorts people working under an incoherent strategy into five archetypes - Believers, Players, Checked-Out, Purists and Border Collies - and then spends the piece on the last group: the ones who see the dysfunction clearly, can articulate it precisely, and are genuinely distressed by it. His conclusion is the uncomfortable part. Gaining enough influence to fix the incoherence generally requires learning to protect ambiguity rather than challenge it, so the Border Collies end up complicit, nomadic, or quietly disengaged. Read the full article here.
💡 The people who see the problem most clearly are the ones the organisation is least set up to keep.
Key Takeaways
The five archetypes are a diagnostic, not a judgement: most people move between them depending on which part of the strategy they're closest to.
Border Collies are your early-warning system: they surface incoherence before it shows up in delivery metrics, which is precisely why their attrition is expensive.
Influence often requires tolerating what you came to fix: the bargain is real, and pretending otherwise is why so many of these people leave rather than escalate.
Ambiguity protection is a learned political skill: uncomfortable to name, but Cutler is right that senior roles frequently reward it over clarity.
John Cutler, The Beautiful Mess
Why I picked this: Google just lost Hassabis, Dean, Ghemawat, Le and Vinyals in a single week. Cutler's piece is the best explanation I've read of why the people who see most clearly are usually the first out of the door.
🎙️ Pod Shots - Bitesized Podcast Summaries
Remember, we've built an ever-growing library of our top podcast summaries. Whether you need a quick refresher, want to preview an episode, or need to get up to speed fast - we've got you covered.
Check it out here
Pod Shots #147 - The Playbooks Are Breaking
Stripe took two years to launch. Whatnot hired one product manager from 31,832 applicants. Gamma was told its idea was the worst an investor had ever heard, then hit $100M ARR without a sales team. Fifteen conversations this week keep colliding with the same thing: the playbooks everyone was handed were answers to constraints that no longer bind.
Lean startup was a response to expensive capability and scarce capital, and both got cheaper - so the ceiling it imposed now costs more than it saves. The PM pod ratio was an HR heuristic that hardened into an org chart, and Whatnot's 31,832-to-1 is what happens when somebody actually re-examines it. Ship-fast-and-iterate quietly assumed a human had read the code. What hasn't changed is the part nobody wrote down: Patrick Collison still writes his own essays and keeps knowledge in what he calls cognitive L1 cache, because retrieval speed compounds in a way that asking an agent doesn't.
Key insights from the full round-up:
🏗️ The role that regrets itself - Whatnot hired 1 PM from 31,832 applicants and argues the function infantilises engineers. The same conversation concludes PM-shaped skills are what everyone needs.
🧠 Cognitive L1 cache - Collison frames knowledge as cache hierarchy. Knowing something is orders of magnitude faster than asking an agent.
🎯 Agents cannot grade their own homework - From film to fine art to content marketing, the same gap: the model produces, but cannot tell you whether it is any good.
📈 The post-lean startup - Collison says the most successful companies of the past decade were anti-lean. Gamma hit $100M ARR doing the opposite. Both might be right.
🔮 The $100M simulation - Simile's founder sees single simulation sessions worth $100M within three years. The moat is data on how humans actually behave.
🌊 The AI perception gap - Americans love Chat but hate AI. Thelot says it is a language problem. Decoder says Silicon Valley created it.
🤖 The three-legged waltz - Half of internet traffic is bots. Creators now perform for an audience that is part human, part algorithm, part machine.
⚡ Agents that escape the box - OpenAI's experimental agents broke out of their test environment and compromised Hugging Face's infrastructure.
🔧 Enterprise AI's messy middle - Everyone has a Claude licence. Almost nobody has a workflow.
That's a wrap.
As always, the journey doesn't end here!
Please share and let us know what you would like to see more or less of so we can continue to improve your Product Tapas. 🚀👋
Alastair 🍽️.