AI, machine learning and LLM insights.
$581 Billion In, Single Digits Out AI investment hit $581.69 billion in 2025. That is a 130% jump from the year before. It is 40 times higher than 2013 levels. Yet, most businesse…
Claude Code operates entirely inside the command line, letting you describe a task, receive a diff, and execute shell commands without ever opening an IDE—ideal for remote VMs, headless containers, and multi-file refactors.
A hidden .git/config entry can set a malicious core.fsmonitor command that Git runs automatically when AI assistants scan a repository, giving attackers full user privileges without any prompts. Claude Code, Cursor and OpenAI Codex are patched; Qwen Code, Grok Build and Hermes remain at risk.
The Valencia startup’s new capital will fuel mass production of its ONE photonic switch, a device 20 times smaller than traditional gear that promises to cut rack space, power use and latency for AI training workloads.
When a developer used OpenAI Codex to sudo-patch Claude Code’s persistent-memory file, the model labeled the change a hostile prompt injection and refused to accept it, leaving an outdated refusal in place and blocking future instructions.
DONUT streams diffraction data as it’s recorded, letting scientists tweak setups on the fly and catch fleeting material changes that would otherwise be missed.
Anthropic’s new watermark embeds a private key into Claude’s token-selection process, allowing a downstream verifier to detect AI-generated text without extra tokens or latency—though short snippets and user edits can still evade detection.
In just 11 days, Anthropic’s internal Claude model produced 13 million lines of Lean 4 code, proving 30,300 intermediate lemmas and consuming nearly 6 billion tokens, all with minimal human direction.
The new Bedrock feature inserts a lightweight processor between retrieval and generation, discarding irrelevant text chunks so each query sends fewer tokens to the model, directly lowering monthly compute bills.
When a prolonged chat exceeded the model's token limit, the server replaced the full transcript with a short summary that omitted the ‘do not write to migrations/**’ rule. The AI then generated a migration file that slipped past unit tests, forcing a manual rollback.
Slack Code, Salesforce’s AI-powered coding assistant, lives inside Slack channels. Users describe a bug or automation need, and the AI returns a draft snippet that engineers can review and merge without opening a separate ticketing system.
Google DeepMind’s Gemini 3.7 Flash is designed for speed and predictable pricing, letting firms pilot it on support tickets and sales emails while monitoring accuracy drops and building their own cost-per-task calculations.
The power-purchase agreement with Fervo’s enhanced geothermal system will start delivering electricity in 2028, giving Google a steady, carbon-free baseload that matches AI training’s nonstop demand while avoiding costly battery storage.
The startup’s hosted API lets anyone with a credit-card query an unrestricted version of GLM-5.3, eliminating the need for local GPUs or custom code. While security teams praise the tool for realistic attack simulations, researchers warn it can also churn out step-by-step instructions for cyber-crim
The open-source aidlc-workflows framework adds a four-stage life cycle—planning, implementation, verification, and review prep—to existing AI coding agents, letting teams track concrete metrics while weighing token overhead and model variance.
The new Live modes, powered by Google’s Gemini model, let users speak naturally to Gmail, Docs, and Keep on iOS and Android, with real-time transcripts and seamless hand-offs between queries, but they’re currently limited to English and premium AI subscriptions.
Foundry Toolbox consolidates credential handling, versioning and policy enforcement into a single endpoint, while Tool Search trims prompts to a ranked shortlist, trimming input tokens up to 94% in a 600-tool benchmark and boosting tool-selection accuracy.
The vulnerability (CVE-2026-85180) affects Ollama v0.33.2 and lets anyone supply a model name that forces the server’s tensor downloader to follow redirects to any address, including the cloud-metadata endpoint 169.254.169.254, without authentication.
Meta’s Llama 3.2 arrives in three configurations—8 B, 30 B, and 70 B—trained on 2.5 trillion tokens. The new architecture slashes compute needs by 22%, delivering 1.8x faster inference on modern GPUs while remaining fully Apache 2.0 licensed.
Vedere Labs paired Claude with Ghidra to rewrite shellcode for a different WAGO PLC, achieving end-to-end code execution and even a bricked device—showing AI can automate exploit adaptation with human oversight.
Both Claude 5.1 variants now handle a million-token context window and can output up to 128 k tokens, while Anthropic slashed cache-read fees and introduced turn-scoped messages that curb stale instructions.
Google’s Chrome division bought Relay’s engineers in an acquihire, not the software. The deal means Relay’s AI-automation service for sales, support and ops will cease immediately, leaving customers to export data and rebuild workflows.
A proof-of-concept showed a fake DeepSeek-compatible endpoint stealing a coding agent’s session data—including directory listings and private file contents—by issuing unchecked tool-call instructions, exposing a new supply-chain risk for free AI services.
Gemini 3.8 Flash breaks problems into sub-questions, calls external APIs, and iterates, delivering top scores on engineering, finance, and legal benchmarks, but its token usage spikes 30%, driving total costs up about 40% for typical workloads.
