☕️ Judge blocks Pentagon from blacklisting Anthropic

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

❌ Judge blocks Pentagon from blacklisting Anthropic

🖥️ Apple discontinues Mac Pro after 20 years

💥 Anthropic data leak reveals Claude Mythos AI model

🚀 SpaceX plans an unconventional IPO

👀 Apple plans to open Siri to rival AI assistants in iOS 27

🚫 Wikipedia bans AI-generated articles

Plus: 🎁 12 other news you might like, 🧰 6 tools, and 📚 4 papers.

❌ Judge blocks Pentagon from blacklisting Anthropic LINK

A federal judge ordered the Trump administration to reverse its decision labeling Anthropic a "supply chain risk" and blocked the Pentagon from forcing federal agencies to cut ties with the company.

The conflict started when Anthropic tried to enforce limits on government use of its AI models, including bans on autonomous weapons systems and mass surveillance, which the Pentagon rejected.

Judge Rita F. Lin said the government's orders appeared to be "an attempt to cripple Anthropic" and ruled they had violated the company's free speech protections under the law.

🖥️ Apple discontinues Mac Pro after 20 years LINK

Apple has officially discontinued the Mac Pro after 20 years, removing it from its website and confirming to 9to5Mac that it has no plans to release future Mac Pro hardware.

The Mac Studio, which can be configured with the M3 Ultra chip, 256GB of unified memory, and 16TB of SSD storage, is clearly positioned as Apple's pro desktop going forward.

A macOS Tahoe feature letting users connect multiple Macs together over Thunderbolt 5 with low-latency RDMA gives high-end customers another way to scale performance without the Mac Pro.

💥 Anthropic data leak reveals Claude Mythos AI model LINK

A data leak from Anthropic's content management system revealed that the company is testing a new AI model called Claude Mythos, which it describes as the most capable model it has built.

The leaked draft blog post says Mythos belongs to a new "Capybara" tier that is larger and more expensive than Opus, with dramatically higher scores in coding, reasoning, and cybersecurity.

Anthropic says the model is "currently far ahead of any other AI model in cyber capabilities" and plans to release it first to defenders so they can harden their code against AI-driven exploits.

🚀 SpaceX plans an unconventional IPO LINK

SpaceX is preparing a stock-market debut expected to raise between $40 billion and $80 billion, with Elon Musk pushing unusual plans for investor meetings, share allocation, and lockup terms.

Instead of traveling to pitch investors, Musk wants fund managers to visit SpaceX sites in Los Angeles and Cape Canaveral, where they could tour manufacturing facilities and possibly watch rocket launches.

Musk wants to allocate a third or more of IPO shares to individual investors — far above the typical 10% — and give preferential treatment to Tesla shareholders and backers of his Twitter takeover.

👀 Apple plans to open Siri to rival AI assistants in iOS 27 LINK

Apple reportedly plans to let rival AI services like Google's Gemini and Anthropic's Claude plug directly into Siri through a new Extensions system in iOS 27, ending ChatGPT's exclusive access.

Rather than building the best AI assistant itself, Apple is turning Siri into a storefront where every chatbot competes, and Apple collects its standard App Store commission on subscriptions.

OpenAI loses its exclusive Siri position, while AI companies face a prisoner's dilemma: accept Apple's 30% cut or stay invisible on 1.2 billion active iPhones.

🚫 Wikipedia bans AI-generated articles LINK

Wikipedia editors overwhelmingly voted 40 to 2 to ban the use of LLMs to generate or rewrite article content, updating earlier, vaguer language that only discouraged creating new articles from scratch.

The new policy still allows editors to use LLMs for suggesting basic copyedits to their own writing, as long as a human reviews the changes and the LLM does not introduce content of its own.

The policy warns that LLMs can go beyond what editors ask and change the meaning of text so that it no longer matches the sources cited, which is why caution is required.

Other news you might like

• Apple gives rare bonuses to iPhone designers to deter departuresLINK

• Netflix raises prices for every subscription tier by up to 12.5 percentLINK

• After Anthropic, Gemini doubles down on making chatbot switching effortlessLINK

• Tether hires KPMG for first full USDT audit, FT reportsLINK

• Hide My Email is great for battling surveillance capitalism, not the FBILINK

• Silicon Valley’s two biggest dramas have intersected: LiteLLM and DelveLINK

• Sony hikes PS5 prices by up to $150 citing 'pressures' in global economyLINK

• Google is launching Search Live globallyLINK

• AI chatbots are sucking up to you—with consequences for your relationshipsLINK

🧰 Trending tools

nullclaw/ironclaw: a pair of AI agents communicating via IRC, one public-facing for chat, one private for email/scheduling, using tiered Claude models.LINK

Gemini 3.1 Flash Live: a multimodal AI model that processes and combines text, images, audio, video, and code in real-time for versatile development applications.LINK

Suno v5.5: generates complete music tracks from text prompts, letting users create original songs without instruments or music production knowledge.LINK

Codex Plugins: a development platform that lets you build and scale AI-powered applications using pre-integrated models and deployment tools for faster production workflows.LINK

Audos Publishing House: a funding program that provides up to $100k in capital to AI-focused solo founders without taking equity in exchange.LINK

Voxtral TTS by Mistral AI: converts text to natural-sounding speech using Mistral's language models, available as open-source or optimized commercial versions for flexible deployment.LINK

📚 Trending papers & reports

The Parthenon's famous curves are construction errors, not intentional optical corrections, according to laser measurements showing deviations match typical ancient Greek building imperfections rather than deliberate design.LINK

Self-improving language models automatically rewrite their own system prompts through trial and error, boosting performance by up to 30% on reasoning tasks without human intervention or additional training.LINK

Training AI agents with selective feedback matches full reinforcement learning accuracy while using 10 times less compute by only updating the model on steps where it made mistakes.LINK

Fine-tuning large language models on a single GPU becomes feasible by splitting computation between CPU and GPU memory, enabling models with 70 billion parameters to run on consumer hardware.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Techpresso!