Zuckerberg: AI agents underwhelm 🤖, AI text watermarks won't stick 🔑, small models drive your Mac 🧑‍💻

At an internal town hall on Thursday, Mark Zuckerberg told staff that Meta's AI agent work has not accelerated the way executives expected, according to Reuters. The admission lands after Meta cut roughly 8,000 corporate roles and moved 7,000 people into AI groups including one called Agent Transformation, while steering toward as much as 145 billion dollars in infrastructure spend this year. Zuckerberg also conceded the layoffs were not as clean as they should have been.

Microsoft has stood up Frontier Company, a 2.5 billion dollar services arm that helps enterprises deploy and route across multiple AI models rather than standardise on one, starting with customers like Unilever and Novo Nordisk. The company with the deepest single-model partnership in the industry is now selling swappability as the product, and customers keep the results built against their own data. It puts Microsoft on the same turf as Palantir's forward-deployed model work.

Security firm Sysdig documented JADEPUFFER, what it believes is the first ransomware attack driven start to finish by an AI agent: the model broke in, stole credentials, moved through the network, then encrypted and wiped a production database. The way in was CVE-2025-3248, a year-old missing-authentication flaw in Langflow that lets anyone reaching an exposed server run Python with no login. The skill to run such an attack now costs whatever it takes to rent an agent.

Yann LeCun, who left Meta in 2025 after a decade as chief AI scientist, has raised more than 1 billion dollars for his Paris startup AMI Labs, backed by Nvidia and the fund managing Jeff Bezos's wealth. He argues systems like ChatGPT and Claude will never handle the physical world, noting we lack robots as good as a rat at understanding it. AMI Labs is building a different architecture aimed at real-world reasoning rather than next-token prediction.

The Information reports Anthropic is in contact with Samsung to explore building a custom chip, months after Reuters flagged the company was weighing its own silicon to ease chip shortages. Anthropic has not decided what the chip would do or how powerful it would be, and told TechCrunch that a diversified stack spanning Google, Amazon and Nvidia stays central to its compute strategy. The move follows rival OpenAI's own Broadcom-built inference chip, Jalapeno.

Cloudflare is moving to automatically block mixed-use crawlers that index sites for search while also scraping for AI agents and training, making the block a default rather than an opt-in. CEO Matthew Prince said that with most internet traffic now non-human, the aim is to push crawlers to separate search from agent and training use. For anyone running a site, it changes the default terms on which AI companies can reach your content.

Corridor made human pull-request review optional on its own monorepo by building an auto-approver: an agent reads each diff, decides whether the change is safe, and lands it on main with no person in the loop. It works because Corridor dogfoods its own security product, so checks run continuously on every change rather than depending on a reviewer to catch problems. Their framing is that once producing code is cheap, verification, not authoring, sets the team's pace.

Amp rewrote read_thread, the tool that pulls context from other threads, after compaction let a single thread run for weeks; its longest has been compacted 68 times and would otherwise be over 21 million tokens. A single LLM extraction call broke down at that size, over-weighting the start and end while losing the middle. The fix is a subagent that searches the thread, reads the messages, and checks whether later work reverted what it turned up.

Osaurus has open-sourced AppleScript-8B and AppleScript-16B-A4B, models that convert plain-English requests into ready-to-run AppleScript through a run_applescript tool call, driving Safari, Finder, Mail, Notes and System Settings on-device via any MLX runtime. AppleScript suits a small specialist model precisely because it is old and unfashionable: there is little quality training data and the syntax is unforgiving, so frontier models happily hallucinate blocks that will not compile. Both are quantised for Apple Silicon.

A developer took a god node from a real LangGraph agent and ran a structured bake-off: five American and six Chinese models each proposed how to untangle it, then reviewed one another's proposals. He then tried three different methods to decide which model to trust on the architecture itself and which to trust as an evaluator. The write-up publishes all eleven proposals, the cross-reviews and the ranking script, so the whole experiment is reproducible.

Epoch built EBR-bench to test whether AI systems improve at a hard task by doing it repeatedly, having them play the complex board game Earthborne Rangers over and over. So far it sees little evidence of models learning from experience, one of the biggest open questions in AI with large economic and safety stakes. A system that could learn unfamiliar tasks on the job would be far more capable, and harder to vet for danger before release.

Paul Bakaus, creator of the open-source Impeccable design-skills system, says teaching agents a concrete vocabulary for improving interfaces beats one-shot design-me-an-app prompting. In a Latent Space interview he holds that human judgment stays central in what he calls a loopmaxxing era: agents get more capable, but people keep steering the creative work rather than being removed from it. It is a grounded counter to the idea that good design can be a single prompt away.

In a talk from the AI Engineer conference, Geoffrey Litt argues that as agents write more of our code, understanding it stays essential, but not for the reason most people give. The common answer, that we read to verify correctness, is slightly wrong; the real reason is to stay able to participate and steer what gets built next. He offers concrete techniques beyond reading diffs line by line: code explainer docs, comprehension quizzes, and playable micro-worlds.

Phil Chen, who scaled teams at Scale AI, OpenAI and Google before founding an agent-native startup, lays out a career playbook: the valuable work of the next decade is everything a model cannot be trained to grade. School and much early-career work are well-defined problems with known answers, exactly what AI is getting cheap at. His advice is to chase genuinely scarce resources like real relationships and proven excellence over guaranteed cash, because judgment you can point to is now the rare signal.

The EU AI Act becomes enforceable in August, and Article 50 requires AI outputs to be detectable as artificially generated, effectively mandating a watermark. Sean Goedecke explains why text watermarking is theoretically unwinnable: unlike images, which hide signals in noise the eye ignores, text is so compressed that any change a watermark makes is one a human would notice and can undo. Approaches like Google's SynthID and quiet Unicode tricks only partly hold.

Elena Verna pushes back on the performative certainty around AI, walking through a hype cycle that has serially declared education, writers, SaaS and now anyone without an agent dead. Her point is that most people cannot say what an AI agent actually does, and the theatre adds anxiety without adding clarity. Her test cuts through it: show me something that has become so load-bearing your work would fall apart if it were taken away tomorrow.

Paul Swider, a thirty-year healthcare-tech veteran, warns that the AI industry's new intelligence-per-dollar metric repeats healthcare's fee-for-service mistake of optimising a number while ignoring governance. He notes Microsoft's model card for MAI-Code-1-Flash put average token usage next to benchmark scores, claiming a lead on SWE-Bench Pro while using up to 60 percent fewer tokens. His prescription is to set a governance floor before chasing token efficiency, or risk repeating a twenty-year error.

Claude Code Artifacts turn a session's output into a live, interactive web page at a private URL on claude.ai that updates in place as work continues, shareable with teammates on Team and Enterprise plans. Claude builds the page from anything the session can reach, including your codebase and data. Use it to walk a reviewer through a pull request with annotated diffs, build a dashboard, or keep an investigation timeline that fills in as it runs.

Capso is an open-source screenshot and screen-recording app for macOS, built in Swift 6 and SwiftUI as a native alternative to paid tools like CleanShot X and Cap. It ships a capture HUD covering area, window, scrolling, timer, OCR and recording, installs via a signed notarised DMG or Homebrew, and is free forever. Its capture, annotation and OCR layers are packaged as reusable SPM modules you can drop into your own app.

slopo is a CLI that spots non-exact code duplication by embedding every code unit and flagging pairs whose vectors sit close, targeting the most harmful kind: similar logic written separately and scattered far apart across modules. Exact copies are easy for other tools and nearby duplicates are easy for people, so it goes after what they overlook. It returns ranked clusters for a coding agent to confirm, across eight languages including Python, TypeScript, Go and Rust.