Claude is AGI

Claude keeps beating Anthropic’s take-home

Claude finished the test faster than most applicants.

Anthropic has used a two-hour performance engineering take-home since 2024, with over 1,000 candidates completing it. Each new Claude release kept solving it, until Opus 4.5 matched the best human scores under time pressure, around 1,579 cycles, while humans needed careful steering and luck.

You could hear the laptop fan spin up during internal tests. What stands out is not that Claude is good at optimization, but that it is now good enough to erase time pressure as a differentiator, which is what most take-homes quietly rely on. The honest admission here is refreshing: banning AI would miss the point, and raising the bar just rewards whoever lets the model run longest.

This is uncomfortable. Anthropic’s solution was to go stranger, using constrained puzzle-like problems that feel more like Zachtronics than real work, trading realism for signal. That feels like a stopgap, not a destination.

If realism no longer works in two hours, what does hiring look like next year?

Apple finally lets Siri become a chatbot

Apple is giving up the argument it kept making.

Apple plans to turn Siri into a full AI chatbot, internally called Campos, likely debuting in iOS 27 and positioned as a centerpiece of WWDC in June. The new Siri would handle both voice and text, looking much closer to ChatGPT than the assistant Apple has defended for years.

You can imagine this being tested late at night on an iPhone screen. What matters here is the reversal. Craig Federighi publicly resisted the chatbot framing, but pressure from ChatGPT, Gemini, and now OpenAI’s hardware ambitions seems to have forced Apple’s hand. This reads less like bold vision and more like Apple admitting its original plan stalled.

That is rare for Apple. If Siri becomes a chatbot, Apple enters the same comparison game it avoided, where response quality and speed are obvious. Choosing Google’s Gemini as a backend partner also means Apple is betting that integration and distribution can outweigh model ownership.

The question is whether users see this as Apple catching up or finally giving in.

Microsoft brings Phi into robots

The robot hesitates, then a human nudges it back on track.

Microsoft Research announced Rho-alpha, its first robotics model derived from the Phi vision-language family. The model translates natural language into control signals for dual-arm robots, combining vision, language, and tactile sensing, and is being tested on platforms like UR5e arms and humanoid robots using benchmarks such as BusyBox.

The demos run at real-time speed. This is Microsoft saying physical AI is no longer just a research curiosity, but a product direction tied to Foundry, Azure, and early access programs. What’s interesting is the emphasis on tactile data and human correction, which quietly admits current robots still fail in messy moments and need help to recover.

That honesty makes the work feel more grounded. If this approach scales, it shifts robotics from hand-coded behaviors toward systems that improve during deployment, using simulation and feedback instead of endless teleoperation. It also tightens Microsoft’s grip on the stack, from synthetic data in Isaac Sim to cloud-hosted models.

The open question is whether adaptability will arrive before trust does.

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.