Alex P @TokenDepth

@emollick Were you able to reproduce this reliably? It does seem quite interesting, possibly explained by different system prompts for different user bases (regardless of the language the query was in).

@emollick Clear reasoning about top tier frontier model policy is important to help any consumer understand what they can expect in the future.

@jyothiwrites Why not both cost per task and by tokens?

@ZenkoZeee @yuhasbeentaken Absolutely, I’m hoping to have an early findings paper with my comparison study out this weekend. @yuhasbeentaken let me know if you’d be interested in having a look.

Hey AI fans: Fable is out (consumer grade Claude Mythos). You get to use it with your Claude subscription until June 22. After that it switches to usage credits.

Anthropic just launched Opus 4.8 and I am already using it. Have you tried it yet? What’s your experience been?

community.mindstone.com/annotate/artic…

Spoke at the Mindstone Silicon Valley AI Meetup last Friday. Topic: Advanced Coding with AI. AI sounds certain even when it's wrong. I built a TDD (test-driven development: write failing tests first, then build only the code needed to pass them) skill live that caught exactly this by forcing failing tests first and stopping for my approval before proceeding. Verification is never optional. Thanks to @MindstoneHQ for hosting! Full recording link in the comments. #AI #ClaudeCode

AI can't read the room. It has zero clue if you're calm or pulling your hair out. No sighs, no body language, no facial expressions. Big tech is racing to fix this with outward-facing cameras on wearables. Not to record you, but to give AI the environmental context it's missing. Even if the camera isn't facing you, it can tell if you're in a stressful environment. The next AI leap isn't better text. It's a system that actually sees your world. #AI #TheoryOfMind

AI behaves like a gas. It fills whatever container you put it in. 💨 I discussed this with researcher Sarah Walsh through the lens of Theory of Mind: the same skill that helps you read people also helps you work with AI. The best power users don't just ask questions. They build a strict container first - a "master prompt" with boundaries, tone, and context. That's the actual secret to prompt engineering. #AI #PromptEngineering

Went to AI Dev 26 in SF thinking AI developer productivity approaches were settling down. Left thinking there's way more to explore than I expected. Cursor just introduced background multi-tasking, and I'd be surprised if others don't replicate it quickly. Nobody has figured it all out yet, and good ideas are spreading fast. #AIDev26 #AI

AI agents can now spend your money. If a workflow hits a paywall while you're away, it just pays. No pause or approval. The tech is genuinely exciting, but the spending controls are barely an afterthought right now. We already know how to fix this: - set hard spending limits, - require approval above a threshold, - log every transaction. Basic stuff. We just need to actually do it. #AI #AIAgents

meetup.com/mindstone-sili…

Tonight I'm presenting at the Mindstone Silicon Valley AI Meetup again at Hacker Dojo. This time: Advanced Coding with AI. I'll walk through how I use Claude Code every day and the features that actually changed how I work. Live demo included. If you're in the Bay Area, come through. Event link in comments. ⏱️ 6 PM PT 📍 855 Maude Ave, Mountain View, CA 94043 #AI #ClaudeCode

Walking the floor at AI Dev 26 in SF. Significant shift toward breaking AI. The industry obsessed over generating more output, faster. Now the energy is moving toward catching what AI gets wrong. Generating code is cheap. Knowing when the AI confidently broke something you stopped checking 20 minutes ago? That's the bottleneck. #AIDev26 #AI

In SF today and tomorrow for AI Dev 26, hosted by @AndrewYNg Say hi if you're around! #AIDev26 #AI

Daily shorts: youtube.com/@tokendepth/sh…

Your delivery app knows exactly how much you are willing to overpay. Delivery platforms watch your shopping habits to figure out what you are willing to pay. Two people looking at the exact same item can get different prices. It sounds bleak, but the upside is that algorithms rely entirely on your routine to function. Once you know what the system is doing, you can change your habits and outsmart it. #AI #TechNews

@ZenkoZeee For example I tend to write a lot like I’m speaking to a friend. I use words like “so”, “like”, “clinical”, or expressions like “I’m unsure about”. And I use these a lot. So, these are my quirks and AI generally speaks in a more clinical and slightly exaggerated fashion.

Everyone is using AI to write faster. The result: the entire internet is starting to sound like one giant, boring robot. We are stuck in a massive feedback loop where AI just keeps training on other AI content. "Fast and clean" is the new average. Stop letting the algorithm iron out your quirks. This is your biggest competitive advantage right now! #AI #Writing