Claude plugin for coding context, sales teams replaced by agents, and xAI's 2 GW power needs
Yesterday, Unsloth released a tool for 2x faster LLM fine-tuning, using 70% less VRAM. This directly impacts iteration speed and compute spend for builders. Separately, new research explains why LLMs miss basic prompt instructions, a critical lesson for designing agentic systems that actually work.
Analysts expect 2026 to see major AI IPOs, with Anthropic and OpenAI potentially debuting publicly. These could collectively surpass hundreds of prior US offerings. SpaceX is also reportedly in the mix, alongside three Chinese tech giants planning $1B+ Hong Kong IPOs.
Read full story→This article argues that in a post-Singularity future, builders should prioritize contributing to broad prosperity or creating lasting work over accumulating personal wealth. It frames the current AI era as a unique historical moment to shape a positive future for humanity.
xAI is expanding its Colossus supercomputer with a third building, "MACROHARDRR," aiming for nearly two gigawatts of power and over one million GPUs. This multi-billion dollar investment boosts xAI's training capacity but forces them to build a natural gas plant due to massive energy demands.
At 70M, many transformer variants cluster tightly; recipe and data dominate. New research shows a diffusion model, Dhara-70M, sacrifices 1.33% accuracy (on specific benchmarks) for a 3.8x throughput boost and superior factuality (measured by a specific metric).
Gemini 3 Flash (68.5%) destroys every premium model on the Misdirected Attention benchmark, including GPT-5.2 Pro (61.7%) and Opus 4.5 (60.7%). The benchmark tests whether models actually read your prompt or just vibe-match patterns. Even top models only score ~50-68% on simple perturbations like "five dead people vs one living person." Translation: your AI features will break on edge cases more than you think, and paying for reasoning tokens won't fix models that don't comprehend the question.
Instagram CEO Adam Mosseri declared the 'intimate feed' dead, stating users 'can’t necessarily trust what you see anymore' due to AI. This marks a fundamental shift in digital authenticity assumptions, requiring new verification UX.
Imprint details their internal AI agent workflows, sharing methods for context window compaction via summarization, structuring subagents for specific tasks, and using a virtual file system (VFS) for state management. They advocate a 'build to throw away' method for learning.
A new book teaches how to build a deep learning library from scratch, focusing on an autograd engine and layer modules using only NumPy. You'll create a custom library capable of training models like MNIST, simple CNNs, and simple ResNets.
SaaStr founder Jason Lemkin states, 'We replaced our team of 10 Sales Execs with 20 AI Agents, managed by 1.2 humans.' He discusses effective AI tools for sales and predicts most SDRs/BDRs will be obsolete within a year.
Unsloth claims up to 2x faster fine-tuning with ~70% less VRAM on supported runs for popular LLMs like Llama, Gemma, Qwen, DeepSeek, and OpenAI gpt-oss.
Claude-mem is a Claude Code plugin that captures tool calls, diffs, and summaries during coding sessions. It compresses this data using Claude's agent-sdk and injects relevant context into future sessions, creating a persistent memory. Implement filters, allowlists, or redaction for secret hygiene.
NVIDIA's DeepLearningExamples repository offers deep learning scripts. Pick one model, run the reference script on your target GPU, capture throughput and accuracy, and use it as a baseline for optimization and vendor comparisons.
Tasker is a free, open-source desktop agent for browser automation. It records actions or takes plain English descriptions, using AI to adapt to website changes and run locally for privacy. It is still brittle on heavy JS apps, CAPTCHA, 2FA, and anti-bot.