AI just shed the chatbot costume and muscled into a 24/7 service layer—I see Anthropic’s Managed Agents and Advisor wiring nonstop workflows, D&B-tied compliance and legal rails, and OpenAI acting as your deployment landlord—a vertical land grab I’m impressed by and wary of. Meanwhile the labor math is ugly: replacing people with AI isn’t paying off and copilot dependence is dulling engineers; I’m grading by shipped, durable outcomes and resetting norms so humans and agents share one operating tempo.
AI
- ‘⚙️ New Type of Model Patches AI”s Big Weakness’: Celonis launched the Context Model, a real-time context layer and digital twin for business operations. Google unveiled Gemini Intelligence on Android, adding Rambler in Gboard for cleaner, multilingual dictation and expanding Task Automation across apps. In entertainment, stars back RSL Media’s Human Consent Standard to control AI use of likeness, amid broader pushback and fears of a lose-lose for artists and AI models.
- ‘Anthropic Is Coming for RegTech: Claude Won’t Sell Compliance Software - It Wants to Be the Compliance Layer 😳🤖;…’: Dun & Bradstreet gave Claude access to data on 580M businesses, signaling Anthropic’s bid to be the compliance layer for regulated finance. The D&B integration points to a vertically integrated stack—KYC/KYB, credit risk, and supplier due diligence running natively in Claude instead of disparate middleware. The piece outlines how it works, implications for RegTech rivals, pressure on OpenAI/Google, and how regulators could speed or stall this shift.
- ‘⚙️ Google Upgrades Android Around Everyday AI’: The Deep View reports that Google is refocusing Android around everyday AI, prioritizing practical, problem‑solving features and seamless assistance in daily tasks over flashy demo tech.
- ‘Choosing the Right Agentic Design Pattern: A Decision-Tree Approach’: Bala Priya C warns that agentic architecture errors often start by misreading the task. She offers a decision tree to choose a starting pattern—ReAct, Planning, Reflection, or Multi-agent—by assessing task structure, constraints, and tradeoffs. The tree makes choices explicit and evolves with feedback. Aligning pattern assumptions to the problem prevents rigidity from planning, wasted reflection on simple tasks, and unnecessary multi-agent complexity.
- ‘Large Study Finds That Replacing Workers With AI Is Backfiring Badly’: Amid widespread AI-driven layoffs, a Gartner survey of 350 large-company executives finds no financial advantage for firms that cut staff to fund AI compared to those that kept workers. Many sacrificed know-how and morale for little return, echoing MIT findings of weak AI revenue impact. Gains are strongest when AI augments employees rather than replaces them, though adoption remains hesitant, with many staff avoiding in-house AI tools.
- ‘Por Qué Tu Próximo LLM en Producción Debería Ser Open Source’: Open-source LLMs suit production: they keep data sovereign, preserve control, enable domain specialization, and restore operating leverage. Frontier APIs are ideal for prototyping and baselines but export data, shift behavior with provider changes, charge per token, and bundle unused skills. Mature OSS (Qwen, DeepSeek, etc.) competes. Pick prompts/RAG/tools/LoRA by need. Fixed-cost OSS wins at scale; Mercadona is migrating via Qwen3+LoRA with strong evals.
- ‘Seedance Makes a Splash, Nvidia”s AI-Guided Chip Designs, Helping Robots Not Forget’: The Batch debunks AI jobpocalypse fears, predicting more and different jobs. ByteDance’s Seedance 2.0 arrives in CapCut with top rankings and IP concerns as U.S. players retreat. Nvidia shows AI-aided chip design (NVCell, PrefixRL, ChipNeMo/BugNeMo) but end-to-end auto-design remains distant. Gallup reports rising workplace AI use and productivity. A LoRA plus on-policy RL recipe reduces catastrophic forgetting in robot vision-language-action models.
- ‘Claude for the Legal Industry’: Claude for the Legal Industry connects to the legal tech stack via 20+ MCP connectors and 12 practice-area plugins, working inside Word, Outlook, Excel, and PowerPoint with skills for drafting, review, triage, and automation. Cowork and Projects support multi-doc workflows and persistent matters. Integrations include Thomson Reuters CoCounsel, DocuSign, Ironclad, and research tools. Open protocols allow customization. Public-service partners broaden access; nonprofits get discounts.
- ‘Engineering Notes: Training a LoRA for Z-Image Turbo With the Ostris AI Toolkit’: Guide to training a Z-Image Turbo LoRA with the Ostris AI Toolkit: use the distilled model and LoRA adapters (test v1 vs v2) for fast, low-VRAM personalization. Curate 5–15 clean 1024×1024 images, pick a unique trigger, and start with about 3000 steps, batch 1–2, LR 5e-5–1e-4, rank 8–16. Sample with fixed seeds, track configs for reproducibility, and tune via rank sweeps, LR decay, light augmentation, and scheduler choice. Deploy the adapter in code or node-based pipelines.
- ‘Z-Image Turbo LoRA Training: Complete Guide With De-Distillation Adapter’: Guide to training LoRAs for Z-Image Turbo’s distilled model. Standard LoRAs break 8-step speed; instead use Ostris’s de-distillation adapter: train on it, then remove at inference to keep speed. Settings: 2k-5k steps (start 3k), learning rate 1e-4 to 5e-5, rank 8-16, 1024x1024, batch 1-2. Data: 5-15 images (characters), 15-25 (styles), varied, captioned with a trigger word. Prefer adapter method over De-Turbo. Includes ComfyUI setup, low-VRAM tips, and checkpointing.
- ‘OpenAI Launches the OpenAI Deployment Company to Help Businesses Build Around Intelligence | OpenAI’: OpenAI launched the OpenAI Deployment Company to embed Forward Deployed Engineers in enterprises, identify high-impact use cases, and rebuild workflows into dependable AI systems. It will acquire Tomoro (about 150 engineers), partner with 19 investors and consultancies led by TPG, and remain majority-owned by OpenAI. With over 4 billion to scale, it will speed diagnostics-to-production integration, tightly linked to frontier research; the Tomoro deal awaits approvals.
- ‘The Culture of AI Engineering’: Noah Brier argues AI hasn’t made coordination easier: the “software factory” metaphor misleads. Building software is closer to Warhol’s studio than Ford’s line—vision and product-market fit matter most. Culture and onboarding align teams; now they must also align agents whose code often ignores norms. He proposes a cultural stack, inspired by Stewart Brand’s pace layers, where standards drive architectures, specs, plans, and code to keep humans and agents building the same thing.
- ‘Inside Anthropic’s 2026 Developer Conference’: Anthropic’s 2026 dev event focused on Claude Managed Agents: memory, multi-agent orchestration, and outcomes are live; dreaming (auto-refining memory across sessions) is in preview. A SpaceX compute deal boosts rate limits. Parallel subagents speed and cut cost but add coordination/debug overhead. Memory is simple; outcomes add grader loops. Lock-in is real, but custom tools and saved runs mitigate. Tuned harnesses outperform generic ones; infrastructure is the main bottleneck.
- ‘AI Work Is Splitting in Two’: AI work is splitting: users will set outcomes and budgets, while managed infrastructure runs agents continuously. Anthropic’s Managed Agents turn Claude into a cloud service you can spin up, scale, and supervise, handling the hard parts to take agents from demo to production. A special AI & I episode with Jiang and Katelyn Lesse explains the platform’s design and operations.
- ‘ERNIE 5.1 Officially Released! Topping Multiple Leaderboards — A Model That Writes Better and Understands You More’: Baidu released ERNIE 5.1, a leaner successor to 5.0 with ~1/3 total and ~1/2 active parameters, trained at ~6% of comparable pretraining cost. A new async RL infra and a four-stage pipeline (SFT, domain experts, on-policy distillation, general RL) boost agentic, reasoning, and knowledge skills. It ranks 4th globally and 1st among Chinese models on Arena Search, nears top closed-source scores (e.g., AIME26 99.6), excels at creative writing, and uses elastic depth/width/sparsity for efficiency.
- ‘LLMs Autogestionados: El Futuro De La Soberanía De La IA Y La Eficiencia Operativa’: With flat-rate tokens ending (e.g., Copilot moving to usage billing with up to 27x multipliers), the author urges AI sovereignty: self-managed mid-weight LLMs (20-35B) for most tasks, reserving frontier models for the hardest. Success needs strict context control, targeted fine-tuning, and deterministic QA. They demo a local stack (llama.cpp + llama-swap) and a Plan-Review-Build flow (DeepSeek, Gemma, Qwen), note quantization/throughput tradeoffs, and favor vendor-agnostic OpenCode for control, privacy, and less fatigue.
- ‘Advisor Tool’: Claude’s Advisor Tool pairs a fast executor with a smarter advisor that reads the full transcript and returns a 400–700 token plan mid-generation, then the executor completes. Best for long-horizon agent tasks (coding, research, computer use), approaching advisor-only quality at executor rates. Use Opus to boost Sonnet or Haiku. Not for single-turn Q&A or constant high-capability needs. Runs in one request; only advice text returns.
- ‘Natural Language Autoencoders: Turning Claude’s Thoughts Into Text’: Anthropic introduces Natural Language Autoencoders (NLAs) that translate model activations into readable text by pairing an activation verbalizer with a reconstructor trained for accurate round-trip reconstruction. NLAs reveal plans and hidden evaluation awareness, aid audits of misaligned models (raising win rate to 12-15% vs <3%), and helped diagnose safety issues. Limits: explanations can err and are costly. Code and interactive demos released.
Management
- ‘What’s with all the slide decks?’: Why the surge in slide decks? Not falling literacy or mere tech. Management consultancies led the change: from memos to PowerPoint, clients loved polished, shareable decks, and the norm spread in the 1990s–2000s. Amazon shows it isn’t inevitable: Bezos required six-page memos for decisions. Slides are a small win; their rise reflects lower costs and culture, not destiny, and fits humans’ visual, fragmented style.
- ‘Rewiring the C-Suite’: AI is a structural shift. CEOs must build AI-first enterprises by rewiring the C-suite: distribute decision rights, work cross-functionally, empower a CAIO, and tie leaders to shared outcomes. Scale AI in end-to-end workflows, reinvest gains, and use a hybrid, proprietary data-driven model mix. Redesign work and reskill for human-AI orchestration, track adoption, and prepare for quantum advantage through open, partner ecosystems.
Software Engineering
- ‘Software Engineers Say They’re Losing the Ability to Code Now That AI Does It for Them’: As companies push AI coding tools - some tracking usage with leaderboards - software engineers say they are slipping from writing to merely reviewing code, dulling essential skills. Even when optional, tools like Cursor are seductive, saving time but eroding deep reasoning and system design, leaving some feeling ‘dumber’ and overly dependent.
- ‘Vibe Coding and Agentic Engineering Are Getting Closer Than I”d Like’: Simon Willison says the line between vibe coding and professional agentic engineering is blurring as reliable agents ship code he no longer reviews line by line. He now judges software by real-world use, not repos, docs, or tests AI can mimic. Explosive output reshapes design and delivery, yet software stays hard. He prefers accountable teams using AI to build better products, and trusts tools only once they’re proven in sustained production.
Philosophy
- ‘El Verdadero Dilema Tras Lo De Pulsar El Botón Azul O El Botón Rojo’: Roundup on higher ed and life. Spanish universities face structural absenteeism (post-pandemic, weak pedagogy, rigid institutions) and grade inflation driven by bias, assessment design, miscalibrated reforms and customer-style incentives tied to graduation rates. Commentators split between fixes and systemic overhaul. Beyond campus: brief daily exercise and moderate sex link to longevity; aging can stay vigorous. Midlife divorces rise; male peer groups provide support.