TLDR AI 2026-02-18

You want AI to be more than a never-ending work in progress, and that means enabling no-code, low-code, and pro-code development without IT gatekeepers standing in the way. Yet that doesn't mean governance goes out the window.

Airia is the enterprise AI platform built to drive AI adoption by unifying innovation and security. Teams at Stryker, BuzzFeed, and ArcelorMittal use it to:

• Test prompts, LLMs, and agent variants in safe, prod-like environments
• End AI anxiety by controlling agent sprawl and implementing automatic guardrails
• Discover risks, orchestrate agents, and monitor threats in one place

Get a demo and make AI adoption a reality

Anthropic has released Claude Sonnet 4.6, upgrading coding, computer-use, planning, long-context reasoning, and knowledge-work performance while keeping Sonnet pricing unchanged. Sonnet 4.6 also introduced a 1M-token context window in beta and is now the default model for Free and Pro users in Claude's apps.

NotebookLM has rolled out Prompt-Based Revisions, a feature that lets users tweak, tailor, and tune slides just by prompting the revisions they want. It currently supports PPTX, with Google Slides support coming soon. A video demo of the feature is available in the thread.

Mistral AI agreed to buy serverless deployment startup Koyeb in its first acquisition, positioning Koyeb's platform and team as a core component of Mistral Compute.

Elon Musk recently went on a podcast to discuss alignment, AI data centers in space, robots, China, and more. Musk is very gung-ho about data centers in space, robots, and making his own fabs. He plans to make virtual humans and robots to turn on an 'infinity money glitch'. Otherwise, China will win - they're already more productive.

OpenAI's acquisition of OpenClaw marks a strategic shift from conversational AI to autonomous agents capable of executing tasks. OpenClaw's popularity stemmed from its unrestrained, robust functionality, combining tool access, sandboxed code execution, and integration with messaging platforms. This move signals a new phase for enterprise AI as companies race to develop secure, deployable versions of dynamic AI agents.

Human judgment is the hardest part to scale. You're relying on people to validate training data, evaluate models, review edge cases, and enforce policies — are they judging consistently and defensibly? Welo Data builds systems that support human + AI judgment at scale. Talk to an expert

ERL trains policies with an explicit attempt → feedback → reflection → revised attempt loop, then reinforced the successful revision back into the base model. The approach improves sparse-reward learning and tool-using reasoning performance while keeping deployment-time inference cost unchanged.

Cohere Labs released TinyAya-Base (3.35B) and instruction-tuned TinyAya-Global plus regional variants, aiming for balanced quality across ~67 languages on consumer hardware. The drop also included a multilingual fine-tuning dataset, new benchmarks, and a technical report for reproducible multilingual experimentation.

WebWorld uses a pipeline of 1M+ open-web interactions to simulate long-horizon (30+ step) browsing tasks, paired with a multi-metric WebWorld-Bench for intrinsic evaluation. Trajectories synthesized from the simulator boosted downstream web-agent performance and transferred to other domains like code, GUI, and games.

Figma users can now bring work from Claude Code into the platform. They just need to install the Figma MCP, type 'Send this to Figma', and the browser's rendered state will be automatically translated to fully editable Figma layers. Workflows are changing, and it's easy to get lost in the momentum of creation. The new Claude Code to Figma Design integration aims to help designers escape that tunnel vision, zoom out, and explore the big picture.

Cursor added plugin support so agents could connect to external tools and extend capabilities via packaged MCP servers, skills, subagents, rules, and hooks.

Meta plans to use millions of Nvidia chips in its AI data centers. The companies recently announced an expanded partnership to deliver personal superintelligence to everyone in the world. The financial terms of the deal have not been made public. Meta plans to spend up to $135 billion on AI this year.

Benign outcomes from technological transitions have always been the product of deliberate institutional design: labor law, antitrust enforcement, public education, and social insurance.

Domain expertise is appreciating in value because AI can't easily replace it.

Three things to consider when deciding what AI to use: Models, Apps, and Harnesses.

Simile, Anthropic, Runway, Goodfire, Fundamental, ElevenLabs, PaleBlueDot AI, Decagon, Flapping Airplanes, Baseten, Inferact, OpenEvidence, humans&, SkildAI, Deepgram, Arena, and xAI.