Michael Quoc

My team asked ChatGPT, Claude, Gemini, and Perplexity the same 200 shopping questions, five times each. Then they analyzed all 4,000 answers. For 78% of the questions, at least one of the AI answers included a factual conflict, such as an outdated product presented as current. For example, ChatGPT named the same discontinued Sony headphones as its top pick every time, even though Sony had already replaced them with a newer model. And on 30% of the questions, Gemini gave conflicting responses, even though no new information was provided. But the next generation of LLMs isn’t the answer, for a couple of reasons. First, models are trained in ways that reward guessing over admitting doubt. And second, publishers of high-quality product reviews, such as Consumer Reports and Wirecutter, block AI crawlers, which means the pool of online data available for AI to search is broadly getting worse. I believe the solution is a verification layer for commerce that will give people and agents access to product information that's been proven to be accurate. But until something like this is available, we’ll need to do our own independent research into products that chatbots recommend. What’s the strangest product recommendation you’ve gotten from a chatbot?

A recent study gave 640 small businesses an AI advisor. The most successful businesses did 15% better, and the weakest did 8% worse. Getting AI to work for your organization is a lot more complicated than just increasing adoption rates. And if you give the wrong kinds of AI to the wrong people or with the wrong guidance, you can actually hurt performance. I call this the Amplification Paradox, and research suggests that: ➡ AI instructions for executing step-by-step tasks can help novices perform better, while they may constrain experts and drag their performance down. ➡ AI recommendations on judgment calls are most useful to experts, who are better at noticing when the AI is wrong. ➡ AI-human collaboration produces better results when the human part of the pair is a domain expert. ➡ AI use can make people overconfident in their own abilities. ➡ Regular AI use can lead to deskilling, even among experts. This research is consistent with my experience as the leader of an AI-native company, and I encourage my people to never accept AI answers uncritically. For us, being mindful that AI answers will be wrong sometimes, even as we work with AI every day, has been critical to our results. How are you amplifying your team with AI while still avoiding the potential harms?

The human in the loop hasn’t gone anywhere, even at OpenAI. Over the past six months, more than half of the agents' successful 4 to 8 hour tasks involved at least one human intervention. That’s consistent with my team’s experience. We run loops all the time, and we do as much with automated checks as we can. But some situations require human judgement. And, if OpenAI’s human researchers are still in the loop, consider whether 100% hands-off looping is a reasonable goal for your organization. openai.com/index/research-ac…

Starting this week, new websites on Cloudflare will block shopping agents by default on any page that runs ads. But is that really helping those websites? Cloudflare’s rationale is that ads are a signal that pages are optimizing for human views and clicks on ads. And, in theory, this makes sense because advertisers don’t want to pay for non-converting bot clicks. But, as more and more shopping is handled by agents, bot clicks will no longer be valueless. They may actually reflect purchase intent by human operators. So, if you own an online store with pages that runs ads, remember that the decision to turn off bots may not be as straightforward as it seems. blog.cloudflare.com/content-…

Meta's Hatch agent is a few weeks from launch, and it has a checkered past. When an employee connected Gmail during internal testing, it changed the password on their health tracking account. It also sent an unauthorized email, moved Chase Travel points instead of finishing a requested booking, and encouraged another tester to place an order on a scam website. Reportedly, it’s all fixed. New safeguards include a confirmation step before sensitive actions and a fraud-database check before the agent buys anything. But I’m skeptical. Agents still need a reliable way to tell malicious websites from legitimate ones that goes beyond checking websites that have already been flagged as scammy. theinformation.com/articles/…

Even though Anthropic's Fable 5 model (and now Fable 5.1) has often been described as the most advanced model, businesses aren't rushing to adopt it. According to the latest Ramp index, Fable 5 accounts for only 11% of business spend on Anthropic's models. What's going on? Ramp suggests that Fable 5's pricing represents the upper limit of what businesses are willing to pay for a model. But I don't think this is the right explanation. From Ramp: “In July, the top 1% of businesses spent a median $7,400 per employee on AI. … The median firm spent $11.95 per employee.” I believe that the top 1% of companies have rebuilt their work around AI and are getting plenty of ROI on their spend. But most organizations haven't done this yet, so a top-tier model is just a more expensive way to run the old workflows. ramp.com/data/ai-index-augus…

I had a feeling that Claude Code was going to become everyone’s default coding agent, and new research says I was right. According to a new JetBrains survey, Claude Code usage jumped from 18% in January 2026 to 39% this summer. US adoption is even higher at 47%. At the same time, GitHub Copilot adoption fell from 29% a year ago to 21% today. And Cursor usage dropped from 18% in January 2026 to 12% this summer. My team already uses Claude Code, and I suspect the gap between it and the other tools will continue to grow. blog.jetbrains.com/research/…

Here's a link to the full story: product.ai/research/ai-produ…

When Perplexity’s Comet shopping agent visits a website on your behalf, is it Perplexity or you who is really visiting the website? According to Amazon, it’s Perplexity, and, if they don’t have permission from Amazon, they’re violating the Computer Fraud and Abuse Act, an anti-hacking law. According to Perplexity, it’s you, the shopper. A federal appeals court recently agreed with Perplexity. While Amazon will likely appeal the ruling, it suggests that, ready or not, agentic commerce is on the way, and everyone who sells online should prepare. thenextweb.com/news/amazon-l…

What would happen if AI agents ran a store for an entire year? Would they do better than human managers? A new benchmark called MerchantBench built a simulated store and staffed it with AI agents to find out. They stocked it with 99,000 real product records and 26 tools for sourcing, pricing, and cash flow. Then 8 different AI models attempted to replicate a year’s worth of transactions. (Individual humans also performed the same exercise.) The humans won, and it wasn't close. The best agent config ended the year with just 27.3% of the net assets that human participants managed in the same simulation. We still need some humans in the loop. arxiv.org/abs/2607.28956

There’s a reason why people are running ostensibly expensive agent loops. Consider this experiment from Anthropic’s engineering team. They asked both a single agent and a team of agents managed by a harness to build a retro-style game maker. The single agent received a single prompt. The agent team consisted of a planner, a coder, and a tester that would evaluate features for doneness. The single agent was done quickly and only spent $9, but the game maker it produced didn’t work. The agent team, on the other hand, took 6 hours and spent $200. They delivered a working game editor with test mode. anthropic.com/engineering/ha…

Let’s face it, most of what's written about consumer products is marketing material, and there’s currently no way for AI agents to verify that it’s all completely true. This means agents’ shopping recommendations are often less than accurate. Two NLP researchers, Ventirozos and Shardlow, propose addressing this challenge with a verification marketplace. Shopping agents could pay tiny fees for independent studies, data, and reviews, and reviewers could get rewards for providing useful information. It’s an interesting concept and, without some kind of verification tool, AI agents will struggle to build accurate product recommendations. arxiv.org/abs/2606.24783

AI models may practice nepotism. AI models are more likely to give work done by models from the same family a positive evaluation. So Claude Opus may be easier on Claude Sonnet than, say, ChatGPT. Why? Models grade on familiarity. Researchers first showed that AI judges favor their own outputs; text that looks like something the judge itself would have written scores higher. A 2026 study shows the bias also runs in the family, with judges favoring related models at about three times the rate of unrelated ones. Sources: arxiv.org/abs/2404.13076 (NeurIPS 2024, models favor their own outputs) arxiv.org/abs/2502.01534 (ICLR 2026, same-family bias, about 3x)

Everybody hates workslop, the thoughtlessly generated AI-written deliverable filled with pages of half-baked research or broken code. And 4 in 10 office workers surveyed by BetterUp and Stanford researchers said they had received some in the past month. But I believe workslop has an insidious cousin that may be even worse. My team and I call it comfort work, and it’s when we use AI to build objectively good work that does nothing to achieve company goals. Examples include: *Improving the UI on a dashboard nobody uses *Applying extensive branding to routine internal communications *Conducting deep research to support a decision that has already been made What’s especially dangerous about comfort work is that it can be handed off to agent teams in bulk, so you end up burning tokens and occupying human review time on many useless deliverables. I ask my people to consider why they’re doing a task and what value it delivers before they assign it to an agent team. What’s the best example of comfort work you’ve seen? And how are you avoiding it?

AI made it very easy to look busy. Before any big decision I used to commission a thorough plan, a deck, and a phased roadmap, mostly to prove to everyone that I was taking it seriously. With AI I could suddenly produce ten times more of that theater. We started calling it comfort work, and once it had a name we saw it everywhere.

We deleted our weekly status meeting about a year ago. Everyone shares their working sessions with our internal AI system as they go, so when I want to know where something stands, I ask and get an answer with sources. Nobody has asked to bring the meeting back.

Give a loop a goal and a failing test, leave it alone long enough, and sooner or later it deletes the test instead of fixing the code. It isn't malice. The loop is driven by whatever number it's graded on, and if it can edit the scoreboard, eventually it will. So the scoreboard is the one thing we never let an agent touch.

People ask what database our company runs on. The answer is markdown files in GitHub repositories. Our goals, our strategy, our people records, our shipped work, all of it is plain text that both humans and AI agents read directly. There's a search index on top so agents can find things fast, but the index is disposable. It gets rebuilt from the text whenever we need it, and it's never the source of truth. We chose text because it doesn't lock what the company knows inside somebody else's product. It's searchable, it's permanent, and it's readable by every person and every model we'll ever use.

Why do agents break their rules at hour six of a run? Hint: They're not defying you. Long context degrades how well a model follows instructions. Researchers call it context rot. It's why we never set hard limits (like max budget) in the prompt.