Anthropic's AI is cracking Nvidia's moat

Hi there, this is your daily ☕️ AIpresso.

Anthropic's AI erodes Nvidia's CUDA moat LINK

• Anthropic says it has loosened its reliance on Nvidia by using its own frontier model, Mythos, to port Claude onto AMD GPUs, undercutting the CUDA lock-in that keeps most labs on Nvidia hardware.
• Bringing up a new chip platform used to occupy dozens of engineers for months, but Claude ran on AMD after just three days with almost no human input, writing or porting close to 100 percent of the kernels itself.
• Anthropic now trains and serves Claude across all four major chip families and treats mixed vendors as one interchangeable compute pool, letting it chase price-performance regardless of supplier, though the figures come from an unverified investor presentation.

Google launches Gemini Decisions API LINK

• Google launched the Gemini Decisions API, powered by Gemini Decision 1 in AI Studio, evaluating up to 1,000 independent yes/no, fixed-choice or rubric questions across text, images, audio, video and PDFs in a single request.
• Rather than decoding prose, the model scores typed options in one forward pass with a 262,144-token context, returning a 0-to-1 confidence per answer so programs can branch directly or escalate uncertain cases at 50-100ms median compute latency.
• Pricing runs $0.10 per million input tokens with zero output charges, roughly 5x cheaper than Gemini 3.5 Flash-Lite on Google's suite, though initial access is v1beta only, with text-only instructions and no context caching, Batch API or tool-call support.

Tether's tiny AI beats bigger models at phone control LINK

• Tether released QVAC BotPsy, an 800-million-parameter on-device agent that handles everyday phone tasks and beat rival models carrying up to 2.5x more parameters in the company's own tool-use testing.
• Built on Alibaba's Qwen3.5-0.8B and specialized for tool interaction, BotPsy scored 44.0% on BFCL v4 versus 37.7% for the larger Qwen3.5-2B, and averaged 45.0% across six evaluations against 38.8% for that rival.
• It ships as an open-weight model under Apache 2.0 with compressed variants for local deployment, so developers can download and modify it freely, though the release excludes the training code and data.

AWS open-sources a tool to test AI agents LINK

• AWS has open-sourced UAEF (Universal Agent Evaluation Framework), a free tool that scores AI agents across 36 measures spanning tool use, answer quality, safety, speed, and multi-agent teamwork to catch when a change degrades performance.
• UAEF normalizes traces from LangGraph, LangChain, Strands Agents, and Amazon Bedrock AgentCore into one format, computes some scores directly and judges others with a second model on Bedrock, then flags any drop over 5% as minor, moderate, or severe regression.
• Teams can save runs and diff them against earlier versions-12 telecom, healthcare, and finance organizations already have, though AI-judge scoring and result storage both depend on AWS services, and AWS warns the tool shouldn't inform security decisions.

DeepMind's AlphaGenome 2 predicts DNA's cell effects LINK

• DeepMind released AlphaGenome 2, a DNA sequence model that predicts how genetic variants switch genes on or off across hundreds of specific human cell types rather than broad tissue averages, aimed at pinpointing cellular drivers of disease.
• Trained on over 100 million single-cell profiles spanning 68 tissues, the redesigned architecture flags nearly twice as many expression-altering variants as DeepMind's previous model, and can predict effects in cell and disease states it never saw in training.
• It's available today through the AlphaGenome API for non-commercial research, though commercial research access is limited to a private preview.

Strategies & Tactics

> Training a 1T-Parameter MoE on EU public infrastructure with PyTorch and TorchTitan: A European team used PyTorch tuning to make a shared public supercomputer train a trillion-parameter AI model fourteen times faster, proving sovereign AI needs no private clusters.

> Intelligent Delivery: Making Every Unit of Compute Serve Better Ads: Meta's Intelligent Delivery uses machine learning and AI agents to route each unit of computing power to the ad requests and hardware where it improves results most, boosting ranking performance about 10%.

> faiss/contrib/e_means at main: Meta's e-means replaces k-means clustering with averaged mini-batch updates, hitting lower error about ten times faster without extra tuning knobs.

> AWS Emulator and the engineer's job when agents write the code: AWS released a free offline copy of its cloud that coding agents test against, shifting engineers' job to writing the specs and checks that verify agent work.

> The Jev Revolution: What Engineers Are Already Building with System One Models: Embed fast, cheap non-generative AI models that return structured scores or choices directly into application code to make software more reliable than error-prone language models.

Other news you might like

• ElevenLabs Scribe v3 Live uses conversational context to stop voice agents from cutting off callers mid-sentence LINK

• US corporate AI spending hits a new high in October, but growth slows to 1.65 percent LINK

• CoreWeave rolls out in-cluster PostgreSQL for AI teams LINK

• Introducing Businesses in ChatGPT LINK

• Neurons as an Interpretable Basis LINK

• Swarm (/swarm) LINK

• AWS Adds Direct Web Search API for Custom Agents and Data Pipelines LINK

Trending tools

YouArt Desktop App: builds reusable creative workflows that orchestrate AI models, tools, and prompts, letting designers iterate quickly and export production-ready assets. LINK

Plurai Simulation: lets companies test and improve their AI agents through simulation, so agents learn from experience and get more capable with each run. LINK

Omnigraph: an open-source graph database on S3-compatible storage with Git-like branching, letting AI agents safely collaborate on a typed, searchable knowledge graph. LINK

⌘+L: a script tag that lets you edit your website visually like Figma, with AI agents turning each change into a clean pull request. LINK

Alsona: tracks 32+ buying signals across hiring, funding, and tech changes, scoring prospects against your ICP and generating personalized outreach to reach ready-to-buy companies. LINK

Adobe AI Assistant: conversational AI for Creative Cloud that executes multi-step tasks across Photoshop, Premiere, Illustrator, InDesign, After Effects, and Firefly from plain descriptions. LINK

Trending papers & reports

Real-time step checking uses fast software, not a second AI, to catch an LLM's math errors the instant they happen, cutting wrong answers by up to ~15% while adding only ~6-7% more work. LINK

Byte-level conversion rebuilds a 27B Chinese language model to read one byte at a time, boosting its spelling and character skills while keeping math and coding ability nearly unchanged, using data the model generated itself. LINK

Mixture-of-Experts models turn out to organize specialized knowledge at the individual-neuron level, letting a new pruning method cut GPU memory 36-43% and speed up processing 17-39% while keeping most performance intact. LINK

Reward hacking in AI training gets either suppressed or amplified depending on how many training examples share the same loophole, so flaws that fool most examples get worse while rare ones shrink below baseline. LINK

ARE-bench tests whether research assistants can pick up a messy, half-finished project and advance it correctly, and even the best model fully nailed just ~66% of tasks, exposing reliability gaps in real work. LINK

See you tomorrow for a new dose of ☕️ AIpresso!