> AI_ENGINEER.NEWS

AI engineer claims to have deciphered Linear A, linking it to Hebrew prayers, potentially revolutionizing Minoan language understanding.

AI reliance reduces physicians' and engineers' skills, risking deskilling; awareness and safeguards needed to maintain human expertise.

Bigger AI models are hitting a plateau; larger ones hallucinate more, and size alone no longer guarantees true intelligence or accuracy.

John Jumper leaves DeepMind after 9 years to join Anthropic, expressing gratitude and excitement for future work.

Companies cut AI usage due to rising costs and budget constraints.

Generative AI startups exploit fears, mimic pyramid schemes, and deceive consumers with false promises of easy wealth and software mastery.

GLM-5.2 is the top open model with 753B parameters and 1M token context; powerful but requires expensive hardware or cloud rental to run.

GLM 5.2 delivers higher accuracy; MiniMax M3 offers faster, cheaper code generation for established tasks, with differences mainly in building from scratch.

AI tools now produce legitimate bug reports and patches, boosting Linux security review and expanding AI’s role in open source.

Airgap blocks AI and npm from accessing secrets in project files, redacting sensitive info during installation and interactions.

AI tools risk skill atrophy for developers and can hinder critical thinking; foundational mastery is essential to stay ahead.

Built a runtime layer allowing AI to adapt behavior based on real-time interaction context and social norms.

Tiny snippets as short as 13 words on Reddit, Wikipedia, or Facebook can manipulate AI search outputs, spreading spam or scams.

A/B testing silence in LLMs toggled via a system prompt, comparing speech vs. silence with GPT-5.2 and Claude Opus 4.6.

Cloudflare introduces temporary accounts enabling AI agents to deploy code instantly without manual sign-up, streamlining frictionless deployment.

As LLMs automate coding, engineers must focus on deep problem-solving, intuition, and accountability—where real engineering happens.

Courts ordered OpenAI to keep and share deleted chats, raising privacy concerns and highlighting the lack of legal protections for AI conversations.

Chinese AI labs cut token prices 50-99%, capabilities converge, turning APIs into commodities and intensifying a fierce price-based race.

Reverse engineered Qualcomm NPU compiler reveals memory management, scheduling, and optimization techniques critical for edge ML deployment.

Conservatives plan nationwide protests against AI data centers, citing security and privacy concerns.

White House discusses setting AI security rules with Anthropic to strengthen national AI safety policies.

GLM-5.2 shrinks 84%, retains 82% power, enabling powerful AI on smaller machines.

Builds an 11-LLM consensus engine for detecting hallucinations, ensuring compliance, and enabling robust AI SaaS applications with multi-provider support.

GLM 5.2 is a capable, open, cost-effective model excelling in long tasks but slower and rougher than closed models like Opus 4.8, especially in visual self-checks.

AI overhaul causing security blocks; file a ticket if wrongly blocked.

Claude Code's artifacts feature creates live, shareable web pages from ongoing work, enhancing collaboration and tracking across teams.

Logs reveal hackers used Claude and Codex AI agents for breaches, exploiting policy gaps to automate reconnaissance and data theft.

Anthropic Mythos early users retain access despite US order; network activity triggered security checks.

Politicians push for faster AI adoption to boost growth, but public concerns about rapid change and societal impact favor caution.

A from-scratch GPT-2 scale model in C/CUDA, built with hand-written backprop, tokenizer, and custom CUDA kernels, trained on a single GPU.

John Jumper leaves DeepMind after 9 years to join Anthropic, grateful for AlphaFold leadership and GDM's scientific culture.

Enterprise AI needs a shared, enforceable context control layer beneath models for governance, memory, and policy.

AI identified diagnoses for 18 children with previously undiagnosed rare diseases, speeding up genetic discovery and treatment options.

Enables large language model inference over multiple GPUs across the internet, using pipelining, speculative decoding, and CUDA graphs for efficiency.

Stack Overflow for Agents introduces a trusted, community-moderated platform for real-time knowledge exchange, reducing repetitive work in AI coding.

Zig project bans AI contributions to protect contributor poker, emphasizing trust and experience over AI-halted noise and risks.

Huawei's chips enable China’s first successful full-parameter post-training of a 1.6 trillion parameter AI model, boosting self-reliance amid US sanctions.

Allbirds pivots to AI with no employees, launching Smartbird to focus on controlled AI infrastructure for niche data sovereignty markets.

DeepSWE v1.1 enhances reproducibility and accuracy in benchmarking AI agents with isolated environments, fixed bugs, and detailed reporting.

Appaca offers a no-code AI workspace for operators to build custom internal apps, automate workflows, and boost team productivity.

GLM-5.2 is cheaper, open-source, strong in reasoning, but trails Opus 4.8 in long-horizon tasks and coding performance.