Today's Author

Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.

Something unexpected just happened. The US government used national security powers to issue an export control directive targeting two AI models: Fable 5 and Mythos 5.

What does that mean? Export controls are rules that restrict who can access certain technology, usually for national security reasons. This order blocks any foreign national from using these two models, including Anthropic's own employees who are foreign nationals.

The practical result? Anthropic had to shut off Fable 5 and Mythos 5 for everyone, not just foreign users. There is no clean way to verify citizenship at the API level, so the safest move was a full cutoff.

• Fable 5 and Mythos 5: fully disabled
• All other Claude models: still working normally
• Anthropic says this is a misunderstanding and is working to fix it

If your app uses other Claude models, you are fine. If you were using Fable 5 or Mythos 5, you need to swap them out now.

Fake accounts steal tokens, burn inference credits, and cycle through free trials at scale.

WorkOS Radar blocks abuse at the authentication boundary, before a fake account ever reaches your product, using device fingerprinting, email signals, and cross-product intelligence from 500k daily checks.

No rules to write, no lists to maintain.

Moonshot AI just dropped Kimi-K2.7-Code, a new open-source coding model, and it's a meaningful step up from K2.6.

The big story here is efficiency. The model uses 30% fewer reasoning tokens than before, meaning it thinks less but gets more done. Less wasted compute, faster answers.

Here's what actually improved:

• +21.8% on Kimi Code Bench v2, a real-world software engineering benchmark covering 10+ languages
• +31.5% on MLS Bench Lite, which tests whether AI can invent new machine learning methods
• Better at long multi-step tasks, like fixing a bug that requires touching 10 files across a codebase

Under the hood it's a 1 trillion parameter Mixture-of-Experts model, but only 32B parameters activate per request, keeping it fast. Context window is 256K tokens.

You can run it today via the Kimi API, or self-host using vLLM or SGLang. Weights are on HuggingFace under a modified MIT license.

Coding agents gave your team escape velocity – verification is the drag holding you back. The faster you ship, the more bugs slip through flows nobody had time to check.

Join live webinar: QA.tech's CEO breaks down the verification gap and how the best teams close it – validation that keeps pace with development, at any scale.

Someone just shipped a Discord bot that reads your voice chat messages out loud using Tomodachi Life character voices. Yes, those goofy little Mii voices from the Nintendo game. It is exactly as fun as it sounds.

It works like this:

• Use /join to pull the bot into your voice channel
• Use /voice to pick your character voice
• Type in the voice chat and the bot reads it aloud in that voice
• Use /support to reach the creator's Discord

The code is open source on GitHub under the repo ttsmodachi, forked from an existing TTS bot and extended with the Tomodachi voice layer on top.

It is currently pending Discord's approval process, so invites may be limited right now. Once approved, it will be listed on a bot directory. Watch the creator's GitHub for updates on when the invite goes live.

How was today's email?

Awesome Decent Not Great

{"AlphaSignal": "214 Barton Springs Rd, Austin, USA"}