Anthropic just published research revealing something unexpected: Claude appears to do a meaningful portion of its thinking in a small internal workspace that was never deliberately designed. It just appeared during training.

The researchers are calling it J-space. Here is what it actually does:

You can watch how it works or explore a live J-space demo yourself.

The researchers are careful to note this does not tell us whether Claude is conscious or feels anything. But finding an undesigned, brain-like structure that powers higher-level reasoning inside a model is exactly the kind of discovery that makes these questions harder to dismiss. Read the full research here.

The bigger signal: 👉 AI interpretability research is starting to find structures inside models that nobody put there. Understanding what those structures do and how to influence them may matter more than benchmark scores over the next few years.

TLDR is the free daily email that does the reading for you. Each issue summarizes the most interesting stories in startups, tech, and programming, curated by ex-Google and Anthropic engineers, so you get the signal without the noise.

It's one email instead of 12 open tabs. A 5-minute read with your morning coffee, and you walk in already caught up.

Tech is just the start. We also cover AI, marketing, dev, and more, so you can follow the topics that matter to your work.

Free, daily, and read by 7M+ subscribers. Subscribe and get your mornings back.

Subscribe for Free

There is a simple reason why most people give Claude vague instructions. Typing is slow, so they cut corners. They skip details, compress context, and end up with outputs that need heavy editing.

Voice dictation fixes that. When you speak, you naturally give more context. You explain things the way you would brief an actual colleague. And for tools like Claude Cowork and Claude Code, that context is exactly what produces better results.

Here is a practical breakdown of what to use:

Start with what you already have

Both macOS and Windows have built-in voice dictation. On macOS, press the Globe key twice. On Windows 11, press the Windows key plus H.

It works, but has real limitations: a 30-second timeout, basic punctuation, and no cleanup for filler words. You end up adapting your speech to the tool rather than talking naturally.

VoiceInk — best starting point for Mac users

A Mac app that transcribes directly into whatever app you are using. Everything runs locally. Nothing goes to the cloud. One-time cost: $39, no subscription.

Requires Apple Silicon and macOS 14.4 or later. For talking to Claude, it covers everything that matters — AI cleanup, custom vocabulary for terms like "Cowork" and "Claude Code," and around 95% accuracy. At $39 once, it is the easiest recommendation on this list.

Superwhisper — the better option for Windows users

Works across macOS, Windows, and iOS. Runs locally when using local models. Handles meeting recordings and audio file transcription on top of standard dictation. Free tier available, monthly plan at $8, lifetime license around $250.

It takes more setup than VoiceInk but is more configurable. Start with the free version before committing.

Wispr Flow — if you want polished output with zero setup

Cloud-based, produces cleaner text by default - better punctuation, automatic bullet formatting, fewer speech artifacts. Costs $15 per month.

For most Claude workflows, that level of polish is not necessary. Claude handles raw dictated text just fine. Where Wispr Flow earns its cost is if you are dictating emails or documents that go directly to humans without an AI layer in between.

The practical takeaway: Try the built-in option first. If the timeout and punctuation issues slow you down, VoiceInk at $39 is a clean upgrade for Mac users. Windows users should start with Superwhisper's free tier.

Talk to Claude the way you would talk to a smart colleague. The outputs will reflect the difference.

Two significant releases landed this week from Chinese labs, and together they tell a consistent story.

Tencent released Hy3, an open-source model that punches above its size on standard benchmarks. Separately, Meituan fully open-sourced LongCat-2.0, a trillion-parameter coding model that the company claims was trained entirely on 50,000 Chinese chips rather than Nvidia GPUs.

Both models are positioned as cheaper alternatives to US frontier models. Both are open-source. And both arrive as China continues its push to reduce dependence on American silicon.

The Nvidia-free training claim around LongCat-2.0 is the detail worth watching. If accurate, it signals that Chinese labs are not just catching up on model quality — they are building the infrastructure to develop at scale without access to the hardware the US has been restricting.

The bigger signal: 👉 The gap between US and Chinese AI is narrowing on two fronts at once: model performance and hardware independence. Both matter for where the next phase of competition plays out.

A few tools quietly worth exploring:

The next wave of millionaires will be people who figured out how to make AI work for them.

The window to get ahead is still open. But not for long.

Here are 200+ proven ways to make money with AI in 2026.

Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.

Claim your free list

See you next issue.