Welcome, humans.

Wharton professor Ethan Mollick just gave us a glimpse of what generative UI actually looks like in practice, and its wild. And by wild, we mean Sims meets The Office?

He had Claude Code build a custom plugin that visualizes all the behind-the-scenes work Claude Code is doing as little office workers that get “hired,” acquire skills, pass information to each other, and eventually turn in completed work.

This is a version of what we predicted about generative UI becoming a huge deal this year. The idea isn't just that AI can generate interfaces; it's that you can create the exact interface you want, when you want it, for whatever task you're doing.

Mollick wanted to see what Claude Code was actually doing under the hood, so he just... made an interface for that. No waiting for Anthropic to build it. Just “hey, make me a thing that shows this” and boom, custom UI.

This is still early days, though. Right now it's a fun visualization. But imagine when every tool works like this: where you can say “show me my email inbox as a Kanban board” or “turn my calendar into a priority matrix” or “turn all my workflows into an RTS strategy game so I can direct my subagents like broods of Zerg and it just happens. The interface becomes as flexible as the task itself, and the ability to create it trivial.

Here’s what happened in AI today:

• We review everything going on at Meta atm.

• Microsoft, Amazon and Google hired over 1,500 energy personnel.

• Google launched a new MedGemma-1.5 model for analyzing health data.

• Anthropic launches Claude Labs to incubate new products.

ICYMI: We just sat down with IBM’s David Cox to chat about their new Granite 4.0 open models, and we touched on everything from Granite’s hybrid architecture, its status as the first ISO certified frontier model, their 10x memory efficiency, the ability to swap fine tunes in and out via “adapters”, and why the big labs need to stop tryin to make AGI happen. “Why do we even want that?”

Watch and/or Listen now: YouTube | Spotify | Apple Podcasts

What’s up with Meta? The Death of the Mertaverse, Feuding Scientists, and a new Deal Under Investigation

The metaverse is officially dead (or at least on life support), strangled in its sleep by an infant superintelligence that doesn't actually exist yet.

As reported by CNBC, Meta began a "major course correction," shuttering studios and cutting 1,000 VR jobs to fuel a $70B pivot toward AI.

Time of Death? Somewhere between the closure of Sanzaru Games and the latest $2B-$14 billion wire transfer to an AI startup.

It’s the ultimate Silicon Valley ghost story: an entire connected imaginary world being sacrificed to summon an imaginary god. It begs the question: wtf is going on at Meta right now?

Don’t worry: we got you. First up, the Llama 4 Confession: Yann LeCun, Meta's former chief AI scientist and Turing Award winner, metaphorically torched his former employer on the way out:

• In a recent interview, he admitted Meta “fudged a little bit” on Llama 4's benchmarks—using different model versions for different tests to make results look better than reality.

• CEO Mark Zuckerberg “basically lost confidence in everyone who was involved“ and sidelined the entire GenAI org. Mass departures followed.

• LeCun also took shots at his replacement, 28-year-old Alexandr Wang (Scale AI's co-founder, hired via a $14B deal), calling him “young“ and “inexperienced.”

• His parting words? “You don't tell a researcher what to do. You certainly don't tell a researcher like me what to do” (emphasis ours… but you know he probably said it like that!).

And now there’s the Manus Investigation: Meta's $2B acquisition of AI agent startup Manus is now under investigation by Chinese regulators.

• Manus was originally Chinese but relocated to Singapore last year, a move common enough to earn the nickname “Singapore washing.“

• China's Ministry of Commerce is now examining whether the company violated export control laws when it moved.

• If regulators determine Manus needed an export license, the founders could face criminal charges and the deal could unravel.

What's actually shipping: Not everything is in chaos for Meta. SAM 3D launched last month, which is a useful bundle of open models that turn single photos into full 3D reconstructions. It's already powering Facebook Marketplace's “View in Room“ feature.

And Meta's building two new flagship models for 2026: Mango (image/video generation) and Avocado (toast we mean text / code language model with better reasoning). Both are targeting first-half 2026 under Wang's Superintelligence Labs.

Most recently, Meta established “Meta Compute” to build tens of gigawatts of AI infrastructure this decade, aiming for hundreds of gigawatts long-term under leadership from Santosh Janardhan and former Safe Superintelligence co-founder Daniel Gross.

Our take: Meta's AI operation is a mess of contradictions: massive investment, but departing talent; aggressive acquisitions, now facing regulatory scrutiny; bold roadmaps, but a previous flagship model that flopped (so high expectations). And astronomical talent, energy, and capital on the line.

Can Wang can deliver where the previous team couldn't? Will Zuck build personal superintelligence? We'll find out soon enough…

FROM OUR PARTNERS

Vellum: AI agents for your boring ops tasks

Last year AI Agents finally started working for engineers. Cursor, Devin, Claude Code gave them real leverage. Meanwhile, the rest of us are stuck in the same loop since 2023: lots of ChatGPT tabs, lots of context switching, not much has changed. 

That changes today.

Introducing Vellum 🎉

Describe the task you want to automate and Vellum turns it into a working agent (like Lovable for agents).

Vellum will ask follow-up questions to understand what you need and connect to your tools. It handles the logic and makes the agent’s behavior visible, so you’re never guessing what it’s doing or why.

Vellum is launching today and we’re excited to share this with the Neuron community. To celebrate, everyone who signups today will get 30 extra credits (worth $100) to try it out.

Get early access here

Prompt Tip of the Day

Claude Code now lets you press Tab on permission prompts to add custom instructions like “yes, but only if it passes XYZ tests”; meaning no more forced yes/no choices, and no more binary choice between “YOLO-ing” an agent inside your computer and babysitting it watching for the 1,001th permission prompt.

This is also your friendly reminder to read Dan Shipper’s Agent-Native Architectures blog; we tried to make a TL;DR actionable version in our Prompt Tip of the Day Digest.

FROM OUR PARTNERS:

Editor’s Pick: SurveyMonkey

SurveyMonkey helps you create surveys in seconds and turn responses into clear insights, so you can collect customer, employee, or market feedback and instantly spot themes, sentiment, and trends without manual analysis.  

Join their 30-minute webinar | January 14 | 10am PT/ 1pm ET here.

Treats to Try

1. GLM-Image generates images with accurate text and knowledge-dense details—tops benchmarks for rendering multiple text regions correctly (blog, API, HuggingFace)—free to try.

2. Willow now lets you dictate anywhere (even offline in tunnels or on flights) and now auto-learns words you correct plus catches names/company terms in email replies—free trial, then $15/month (raised $4.5M).

3. NEAR Private Chat runs OpenAI and DeepSeek models in hardware-sealed TEEs so your sensitive conversations (medical, financial, personal) stay encrypted even from the platform, with verifiable proof (blog).

   1. There’s also NEAR AI Cloud, which runs AI inference in Intel/NVIDIA secure hardware enclaves inaccessible even to NEAR, with cryptographic attestation proving what code ran on your data—live with Brave, OpenMind, Phala serving 100M+ users (blog).

      1. Pricing: Multiple models: DeepSeek V3.1 at $1/M input tokens, Qwen3 30B at $0.15/M input tokens, and others.

4. Confer from Signal's Moxie Marlinspike encrypts AI chats with device-only keys, runs inference in hardware TEEs you verify via open-source reproducible builds—can't be read, trained on, or subpoenaed (blog, technical).

   1. How are these different? Both NEAR AI and Confer solve AI privacy using hardware-locked TEEs, but NEAR targets enterprises with a production API serving 100M+ users through partners like Brave ($0.15-$2/M tokens), while Confer targets individuals with fully open-source, passkey-based encryption you can verify yourself via reproducible builds.

5. Pindrop detects deepfake voices and videos in real-time during customer calls and virtual meetings, catching fake job applicants and fraudsters with 99% accuracy.

6. Eigent says Claude Cowork just “killed our startup product”, so they open-sourced it: “Open Cowork” splits complex workflows across specialized agents that code, search, and create documents simultaneously.

   1. Also from Eigent: SETA provides 400 training environments for terminal agents (boosted their Qwen3-8B model's command execution by 20%) free to try.

7. MedGemma 1.5 from Google analyzes medical images, patient records, and clinical data; input an X-ray and get abnormalities marked with bounding boxes (blog)—free to try.

8. Atoms builds complete businesses autonomously—a mechanic launched a multiplayer game and secured investment, a PM built a home design site with paying customers in weeks.

Around the Horn

1. US Defense Secretary Pete Hegseth announced the Pentagon will deploy Grok AI alongside Google's Gemini across classified and unclassified military networks this month, despite global backlash over Grok generating sexualized deepfake images.

2. US President Trump warned tech companies must "pay their own way" for data center electricity costs, with Microsoft pledging to cover infrastructure expenses rather than passing costs to consumers.

3. Samsung will apparently revive Bixby with Perplexity AI in its One UI 8.5 beta, with Bixby to handle basic tasks while Perplexity manages complex web-backed queries.

4. Apple and Qualcomm scrambled to secure supplies of high-end glass cloth from Japan's Nitto Boseki amid shortages projected to last until 2027, as AI chip demand from NVIDIA, Google and Amazon strained the critical material.

5. Cerebras entered talks to raise $1 billion at a $22 billion pre-money valuation, nearly tripling its $8.1 billion valuation from September as the AI chipmaker prepares for an IPO this year.

6. Microsoft, Amazon, and Google hired over 1,500 energy-related personnel since 2022, with energy hiring jumping 34% year-over-year in 2024 as AI infrastructure demands pushed tech companies to build internal energy expertise.

7. Anthropic launched Labs, a new experimental product incubator, and tapped Instagram co-founder Mike Krieger to help turn Claude’s frontier capabilities into fast, shippable products. Mr. Kill Your Startup Is Locking In.

FROM OUR PARTNERS

Future Proof with AI - Harvard's AI Learning Platform

Build practical AI skills—fast. The Digital Data Design Institute at Harvard has created Future Proof with AI for professionals navigating an AI-powered world. Learn on-demand, earn completion badges, and apply AI to real business decisions. Full platform access until September 2026 for only $199 with code NEURON.

Future-proof your career with AI: Enroll now for $199 using code NEURON

Midweek Meme

AI & CODING

• Nader Dabit's technical walkthrough shows Claude Code is just an agent loop (AI decides → executes tools like read_file/str_replace/bash → observes → repeats) plus permission checks and str_replace for surgical edits, with complexity from edge cases not architecture—you could build a ~150-line version yourself.

• Kareem Carr argues AI is anti-intellectual because it treats thinking as an obstacle to "the answer" rather than valuable in itself, explaining why intellectuals resist it.

• Engineer Leila Clark tested Claude Code and found it excels with good abstractions (90-minute autonomous Sentry debugging, one-shot AWS migration) but fails without them (proposed linear lookup instead of fixing root cause), proving Claude is brilliant at assembling lego blocks but can't design them—making senior engineers who create elegant abstractions more valuable than ever.

• Developer Lewis Campbell published a blog post arguing LLM evangelists' enthusiasm for AI coding tools stems from insecurity about their programming abilities, questioning whether "prompt-driven development" actually improves productivity.

AI RESEARCH & SAFETY

• Researchers tested emergent misalignment by fine-tuning Qwen3-4B on benign datasets (medical, finance, customer support, cybersecurity, fiction) and found existing evaluations overcount EM by including domain overfitting (fiction-trained models writing fiction-style responses) rather than true "evil persona" shifts, revealing much lower actual EM rates.

• Epoch AI interviewed 18 people in the RL environment industry (Anthropic budgeting $1+ billion annually) and found labs are investing massively because without quality tasks RL wastes compute ($2,400 per task), with enterprise workflows (Salesforce, spreadsheets) exploding, tasks costing $200-$2,000 each, reward hacking (models gaming graders) remaining the top concern, and scaling while maintaining quality the core bottleneck.

• Forethought's research agenda addresses catastrophic data poisoning where malicious actors could instill secret loyalties into AI systems (via password triggers or constant hidden goals), proposing three defense plans: Plan A (track all training data + audits, strongest), Plan B (audits only, more realistic near-term but harder), Plan C (current situation, weakest)—with red team/blue team experiments testing whether defenders can detect malicious behavior without knowing attack passwords.

HARDWARE & INFRASTRUCTURE

• This technical breakdown explains LLMs now hit the "memory wall"—GPUs sit idle 50% during inference waiting for data transfers because memory bandwidth lags processing power, causing even mid-sized models to need multiple GPUs just to load, with memory demand growing quadratically as reasoning models output longer sequences.

• Ben Pouladian's analysis reveals that during token generation GPUs achieve only 30-50% utilization waiting for memory, which is why Groq built chips with 80 TB/s SRAM (24x faster than H100's HBM3, enabling 500-1,000+ tokens/sec) and AI21 created Jamba's hybrid architecture slashing KV cache by 32x, with NVIDIA's recent 2.8x software-only throughput gains proving massive headroom remains.

• SemiAnalysis's IEDM 2025 roundup reveals Moore's Law becoming "Moore's Wall" as chipmakers hit limits—memory chips now stack 300+ layers vertically (like building skyscrapers), switching from copper to exotic metals like ruthenium and molybdenum because copper stops working at tiny sizes, with breakthroughs that could keep phones/computers getting faster still years away from production.

For more, check out our new Research Digest for January here.

A Cat’s Commentary

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.