AI/ML Digest

Morning Edition Thu Jun 11

Anthropic reversed a policy in Claude Fable/Mythos that would have silently limited effectiveness for requests related to frontier LLM development, apologizing for the misstep. Separately, a former xAI engineer sued the company claiming wrongful termination after raising Grok safety concerns, and Amazon secured $17.5B in bank loans days after a $14B Canadian bond sale.

Anthropic reverses silent Claude Fable LLM-research throttling policy

Anthropic will make Fable 5 safeguards visible after outcry over undisclosed limits on frontier LLM development requests.

xAI engineer sues over firing after Grok safety complaints

Devin Kim filed suit in California claiming xAI terminated him for raising safety concerns about Grok's WMD and bias risks.

Amazon borrows $17.5B from banks for AI buildout

Loan from Citigroup, JPMorgan, Wells Fargo and others follows a $14B bond sale, totaling $31.5B in 48 hours.

SIA self-improving agent framework achieves 56.6% LawBench gain

Open-source meta/target/feedback agent loop reports 56.6% LawBench gain and 91.9% runtime reduction on GPU kernels.

Frontier LLMs score 4–8% on black-box vuln detection benchmark

Six frontier models hit only 4–8% ground-truth coverage; domain-specialized agents with structured methodology exceed 50%.

datasette-agent 0.2a0 adds mid-execution user questions and save_query tool

New ask_user() suspends agent turns for human input; save_query tool requires approval before persisting SQL.

SpadeBox: sandboxed file, network, and JS runtime tools for AI agents

Rust library with Python/JS bindings provides cap-std file sandboxing, domain allowlisting, and secret management for agent tools.

AgentsView: local cost tracking across Claude, Forge AI coding agents

Single binary syncs agent sessions into local SQLite and shows daily cost summaries; no accounts required.

Amazon Bedrock team delivers 18-month project in 76 days with AI-native workflow

Six-engineer team shipped more production code in five months than prior ten years; post details three workflow patterns.

PyTorch profiling part 2: fusing nn.Linear layers into an MLP

Hands-on walkthrough of kernel fusion for MLP blocks on A100, with scripts and profiler trace analysis.