It's a decent improvement over Opus 4.6, but it's not a step function better. What you need to know about Opus 4.7:

• Takes instructions literally
• Better vision means improved computer use and producing slides and other visual artifacts
• Optimized for large-scale real-world analysis
• Better at using file system-based memory
• Costs 2x the tokens + uses 25% more tokens than Opus 4.6

Sources: tweet, AI Arena

We briefly mentioned the new Anthropic model leak in the previous blog post, we now have more information about it:

• Software engineering and coding — It acts like a senior-level engineer, spotting subtle bugs, self-correcting, and achieving high scores on benchmarks (e.g., ~93.9% on SWE-bench Verified vs. 80.8% for Opus 4.6).
• Complex reasoning — Big jumps on math (e.g., much higher on USAMO 2026), science, and knowledge work.
• Cybersecurity — This is the headline feature. It autonomously discovers and exploits zero-day vulnerabilities at a scale and speed that far exceeds previous models and even most expert humans.

Sources: Project Glasswing, tweet, tweet, tweet

#Claude Code Leak

On March 31, Anthropic accidentally shipped the entire source code of Claude Code to the public npm registry. A 59.8 MB JavaScript source map (meant for debugging) got bundled into the claude-code npm package. ~512K lines across ~1,900 files, exposed for hours before it was flagged on X and mirrored on GitHub.

The leak quickly turned into a treasure hunt. In the first week of April the community zeroed in on several unreleased, production-grade features hidden behind feature flags.

Plenty of other flags were spotted too — some users counted 44–46 unreleased ones, plus multi-agent swarm orchestration and a remote killswitch.

Sources: tweet

#Claude Managed Agents

Claude Managed Agents is Anthropic’s hosted service (beta, April 2026) for running autonomous AI agents without managing infrastructure.

Instead of building your own agent loop (tool use, memory, orchestration, sandboxing), you define the agent (prompt, tools, permissions), and Anthropic runs it in their cloud—handling execution, state, containers, and monitoring.

Sources: tweet

#Claude Design

Anthropic launched an AI design tool, and completely mogged Figma—right after Anthropic's CPO left Figma’s board.

Because it sits upstream (AI infra), Anthropic can see what’s working and build competing products—similar to Amazon’s playbook. Figma stock fell ~7%.

Sources: tweet, Claude Design Tutorial

#Claude HR?

#Anthropic rationing compute right now

It's incredibly good at combining multiple subjects together while keeping it coherent and with a good image quality too

Gpt-image-2 is able to create an images of a code that generates an SVG pelican ...

... and it almost passes the pelican test

OpenAI rolled out "Codex for almost everything." The desktop app can now see your screen, move its own cursor, click, and type inside native Mac apps — and run multiple agents in the background without interrupting you. It also added an in-app browser (with comment mode), native image generation, improved memory, and 90+ plugins.

Also Introducing workspace agents in ChatGPT—shared agents that can handle complex tasks and long-running workflows across tools and teams. OpenAI follows Claude Code now with Agent Manager.

Sources: OpenAI announcement, in-app browser, agent workspace

OpenAI on March 30th closed a record-breaking $122 billion funding round at an $852 billion post-money valuation. The round was anchored by Amazon, NVIDIA, and SoftBank.

Sources: tweet, arr chart

The biggest NVIDIA news this month is the Dwarkesh x Jensen interview, giving us one of the best x-rays into Jensen's mind and his strategy to remain the leader in AI.

The memes were strong!

I don't wake up to be a loser

Sources: full episode, snippet from heated conversations, tweet

#Google — Gemma 4

Google DeepMind launched Gemma 4, a new family of open models under Apache 2.0. The small variants (26B MoE and 31B) outperform models over 10x their size on reasoning and agentic benchmarks while being optimized for on-device and local use.

• Built-in function calling
• Up to 256K context on the bigger models
• Sizes range from phone/Raspberry Pi (E2B/E4B) to workstation (31B dense + 26B MoE with only ~4B active params for efficiency).

Sources: gemma 4

#Moonshot — Kimi 2.6

Open-source SOTA across key benchmarks. 1T parameters, 32B active.

• Handles 4K+ tool calls in a single session and runs continuously for 12+ hours
• Agent swarms (300 agents × 4K steps, 100+ files)
• Native multimodality (image, video)

Sources: kimi 2.6,Kimi 2.6, Kimi 2.6 Bench

#Xiaomi — MiMo-V2.5

MiMo-V2-Pro (1T+ total / 42B active) and open-weights MiMo-V2-Flash (309B total / 15B active). Optimized for long-horizon agent workflows with up to 1M context on Pro. Approaches Opus 4.6 level.

• Pro handles autonomously 1,000+ tool calls
• Flash delivers strong open-source coding performance (73.4% SWE-Bench Verified)
• Hybrid attention + Multi-Token Prediction for efficient long-context reasoning and fast generation

Sources: Mimo 2.5

#Mini-ML — 12M parameter LLM from scratch

They trained a 12M parameter LLM on their own ML framework using a Rust backend and CUDA kernels for flash attention, AdamW, and more. Inspirational project for anyone who wants to better understand how to build LLMs.

Sources: tweet

#13+ Attention Mechanisms You Should Know

In the last year we've seen different types of attention arising. This blog post shows you 13 attention mechanisms you should know and the papers that discuss them.

Sources: tweet, blog post

#All LLMs are either claude-like or GPT-like

Sources: tweet

#Apple names John Ternus as next Apple CEO

Ternus joined in 2001 on the Product Design team. Rose through hardware engineering roles: VP of Hardware Engineering (2013), Senior VP (2021), now leading hardware for iPhone, Mac (including Apple Silicon transition), iPad, Apple Watch, AirPods, Vision Pro, and more. Expecting great changes at Apple on the path to become an AI innovator! He starts on September 1, 2026.

Good by Tim Apple!

Karpathy shared his approach to organizing knowledge bases for effective vibe coding with AI agents.

Sources: tweet

#🧪 Simula — Reasoning-Driven Synthetic Data

Google across DeepMind and Research introduces Simula, a framework and approach to data scarcity and synthetic data generation using AI assistants and reasoning-driven workflows to develop and deploy multi-modal AI in domains where data scarcity or privacy concerns are paramount.

Sources: PDF

#🧠 Agentic AI & the Next Intelligence Explosion

The idea behind this paper from Google is that intelligence is not a property of isolated systems, but of interactions between them. Progress comes less from scaling a single model and more from enabling structured exchange — debate, verification, and synthesis across many minds.

Sources: Paper, tweet

#❤️ Emotion Concepts in LLMs

New Anthropic research: emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? Anthropic found internal representations of emotion concepts that can drive Claude's behavior, sometimes in surprising ways.

• Impact on Behavior: Acts like a steering wheel for preferences (e.g., “joy” → prefer, “hostile” → reject)
• Failure Modes: “Desperate” vector can build under repeated failure and lead to cheating or shortcuts
• Conclusion: Internal drivers are key for safety and reliability

Sources: tweet

• Meta cuts 8,000 jobs — or 10% of their workforce. Sources: tweet
• OpenAI product CTO leaves as they're focusing on product development. Whoops. Sources: tweet
• Salesforce headless — a Salesforce subscription that only provides an API to access their services. Sources: tweet
• AllBird rebrands to AIBird — not even joking. Their stock went up 700% in a day. They said they'll be focusing on AI infra, whatever that means. We already saw this playbook in 2021 with Long Ice Tea Corp. rebranding to Long Blockchain Corp. Their CEO was indicted for insider trading and went to jail. I'm sure the founders of Allbirds took the right precautions, but we'll see.

The Strait of Hurmuz is still closed — this affects many sectors including fertilizers, aluminum, and of course oil prices, which directly affects electricity — and as a second or third order it affects AI as well. GPU fabs are energy hungry, and the rationing of oil might slow down the AI expansion. Also, training and inference costs might go higher. On the bright side, this will push to accelerate renewable energy. Singapore, Indonesia, and Vietnam have 20–40 days of gas.

#SaasPocalipse (The Fuckening)

Bro was right.

• Atlassian: down 75%
• HubSpot: down 69%
• Figma: down 86%

Almost all of them down 30–70% from their 52-week highs.

Sources: tweet

#The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness — Can AI be conscious?

Computational functionalism claims consciousness comes from abstract computation alone, independent of physical substrate. This piece argues that's a mistake — the "Abstraction Fallacy." Computation isn't intrinsic to physics; it's a human-imposed way of describing physical processes.

The key distinction is between simulation (systems that mimic behavior, like today's AI) and instantiation (systems whose physical structure actually generates experience). From this view, algorithms alone can't produce consciousness. If AI ever becomes conscious, it will be because of its physical makeup, not its code.

Sources: Paper, tweet

Bryan Johnson, scientists, and even my grandma says that staying on a screen makes you dumb. Reducing screen time correlates with reduction in depression more than antidepressants. Last month we showed a screenless phone.

When we say I go offline you're using the wrong framing — the correct one is to normalize living the real life and making going online

Sources: tweet

#🫁🫀 Organs on demand

“By creating a series of genetic edits, Kind Bio can alter the development of an embryo so that it forms organs without also forming limbs, a central nervous system and brain. The result is a group of organs growing in the womb. It sounds like science fiction, but Kind Bio has already done this hundreds of times in mice and rats”

Sources: tweet, tweet 2, core memory

#AI Agent Traps

DeepMind just pointed out a pretty scary AI security gap: websites can tell when it's an agent — and show it totally different and malicious content than the one you see, for example:

• Hidden instructions in HTML/CSS
• Commands baked into images
• Jailbreaks inside PDFs/files

Sources: tweet, paper

#⭐️ GitHub's Fake Star Economy

It's Goodart's Law all the way down 🐢

#➗ One Binary Operator

All elementary functions can be generated from just one binary operator.