Presented by
Good morning, AI enthusiasts. Remember how everyone assumed DeepMind's Demis Hassabis was purely in it for the science? A new book out today reveals he spent years quietly running a high-frequency trading team inside DeepMind to beat Jim Simons. Google found out and shut it down.
Was this just a plan for financial independence, or a sign that even science's most idealistic founders have a second agenda they'd rather keep quiet?
In today's recap:
DeepMind's Hassabis secretly ran a hedge fund
OpenAI model solves three open math problems
Build email triage with Gmail AI Inbox
ZAI launches multimodal coding model GLM-5V-Turbo
4 new AI tools, news, and more
GOOGLE DEEPMIND
Hassabis secretly built a trading desk inside DeepMind
Eleanor Shakespeare
Recaply: A new book just revealed that DeepMind's Demis Hassabis secretly built a 20-person trading team inside the company in 2016, aiming to beat Renaissance Technologies at algorithmic trading before Google shut it down.
Key details:
Hassabis hired about 20 researchers to train trading algorithms, exploring a partnership with BlackRock, with the aim of building a revenue stream independent of Google.
The project targeted Renaissance Technologies, whose Medallion fund is the most successful hedge fund in history. It never made money and was quietly disbanded.
"Rentec operated in secret, which Demis loved," a source told author Sebastian Mallaby, whose new book "The Infinity Machine" reveals the project for the first time.
The trading operation ran through 2016 before Google's wariness ended it. It was never publicly disclosed until Mallaby's book, out now.
Why it matters: Hassabis is known as one of the most idealistic figures in frontier AI. But this story tells a different one. A secret trading team, a BlackRock partnership, and a plan to beat the world's best hedge fund don't fit the usual narrative. It shows that the tension between science and money inside AI labs runs deeper than most people knew.
PRESENTED BY GLADLY
The AI your stack deployed is losing customers.
You shipped it. It works. Tickets are resolving. So why are customers leaving?
Gladly's 2026 Customer Expectations Report uncovered a gap that most CIOs don't see until it's too late: 88% of customers get their issues resolved through AI — but only 22% prefer that company afterward. Resolution without loyalty is just churn on a delay.
The difference isn't the model. It's the architecture. How AI is integrated into the customer journey, what it hands off and when, and whether the system is designed to build relationships or just close tickets.
Download the report to see what consumers actually expect from AI-powered service — and what the data says about the platforms getting it right.
If you're responsible for the infrastructure, you're responsible for the outcome.
ZAI
ZAI releases multimodal coding model GLM-5V-Turbo
ZAI
Recaply: ZAI just launched GLM-5V-Turbo, its first model built to natively handle images, video, and documents alongside code, claiming leading performance on multimodal coding, tool use, and GUI agent benchmarks.
Key details:
GLM-5V-Turbo reads screenshots, design files, and document layouts natively, built for the full loop of "understand the environment, plan actions, execute tasks" in agent workflows.
The model supports 200K context and 128K output tokens, and integrates directly with Claude Code and OpenClaw agent frameworks.
ZAI describes it as "deeply optimized for agent workflows," positioning it as a companion model to agent systems rather than a standalone chat model.
GLM-5V-Turbo is available now at ZAI chat and via the API, with a Coding Plan trial application for team access.
Why it matters: Most models treat vision as an add-on. ZAI built it in from the start. The bet is that the next wave of coding agents won't just read text, they'll read screens, mockups, and design files too. By plugging directly into Claude Code workflows, ZAI is positioning for that future. It's a narrow bet, but a well-timed one.
TUTORIAL
Build an AI-powered email triage system with Gmail AI Inbox
Recaply: In this tutorial, you will learn how to configure Gmail's new AI Inbox to automatically surface your most important to-dos and group low-priority emails into topic clusters, so you spend less time triaging and more time acting.
Step-by-step:
Subscribe to Google AI Ultra ($249.99/month) at one.google.com, which bundles Gmail AI Inbox with Gemini Advanced and other Google AI features.
In Gmail on the web, find "AI Inbox" above your standard "Inbox" in the left sidebar. Click it to open the AI-powered view alongside your regular inbox.
Review "Suggested to-dos" at the top. Gemini 3 pulls action items from your emails: bills due, reminders, and tasks that need a response. Each item links to the source email.
Below your to-dos, browse "Topics to catch up on," where Gmail groups lower-priority emails into clusters like Events, Travel Planning, or Health and Wellness, so you process batches instead of single messages.
Give feedback on surfaced items using the thumbs up/down option. Gemini 3 learns what you flag as urgent vs. low-priority over time.
Pro tip: Keep your standard Inbox open alongside AI Inbox. Use AI Inbox as your daily action layer, and reserve the standard Inbox for an end-of-day sweep. This two-pass system keeps urgent items from slipping through.
AI RESEARCH
OpenAI model solves three open Erdős math problems
OAI/Screenshot
Recaply: OpenAI just published a paper solving three problems posed by mathematician Paul Erdős, with the five co-authors crediting an internal model at OpenAI entirely for the proofs. Their job was only to clean up the writing.
Key details:
The model generated all three proofs on its own. The researchers write that their role was "simply to digest the proofs and modify the write-ups for clarity and elegance." They contributed no math.
The three problems span different areas of mathematics and had been open for decades. They're real research questions, not benchmark tasks designed to be solved.
OpenAI CPO Kevin Weil replied that "not only is AI solving more open problems, its proofs are getting more elegant as the models improve." GPT-5.4 Pro could solve two of the three in under ten attempts. The internal model isn't named.
The paper was posted to arXiv on March 29, 2026, with all five co-authors from OpenAI.
Why it matters: AI math benchmarks involve problems with known answers. Erdős problems don't. They're real open questions that stumped top mathematicians for decades. When OpenAI's CPO says the proofs are getting more elegant as the models improve, that's a signal this isn't a ceiling. The gap between AI test performance and AI doing real research is closing faster than most people expected.
TOOLS
Trending AI Tools
🎨 FAUNA - Flora AI's creative agent
🎤 Atlas 1 - Willow's frontier speech-to-text model
🖼️ Wan2.7-Image - Alibaba Wan's unified model for image generation and editing
⚙️ GrandCode - DeepReinforce's multi-agent competitive programming system
NEWS
What Matters in AI Right Now?
Anthropic issued copyright takedowns removing over 8,000 copies of accidentally exposed Claude Code source instructions from GitHub, later narrowing the request to 96 repositories after initially overreaching. Anthropic called it "a release packaging issue caused by human error."
Google rolled out AI Inbox in Gmail to AI Ultra subscribers, a Gemini 3-powered interface that surfaces to-dos and topic groupings instead of a standard inbox, available to AI Ultra members at $249.99 per month.
OpenAI shares are becoming nearly impossible to sell on the secondary market, with about $600 million in shares being shopped by institutional holders as investors pivot to Anthropic, according to Next Round Capital.
Cognichip raised a $60M Series A led by Seligman Ventures, with Intel CEO Lip-Bu Tan joining the board, bringing total funding to $93M for its platform designed to reduce chip design cycles by 50% and design costs by 75%.
Willow launched Atlas 1, a frontier speech-to-text model built on proprietary human-powered transcription infrastructure, claiming it outperforms ElevenLabs, Deepgram, and OpenAI transcription by a wide margin.
DeepReinforce announced GrandCode, a multi-agent reinforcement learning system that ranked first in three consecutive Codeforces live competitions, beating all human participants including legendary grandmasters in rounds 1087, 1088, and 1089.
Anthropic signed an MOU with the Australian government to cooperate on AI safety research, with CEO Dario Amodei meeting Prime Minister Anthony Albanese, alongside AUD$3M in Claude API credits for four Australian research institutions.
Alibaba's Wan team released Wan2.7-Image, a unified model for image generation and editing with support for realistic face control, multilingual text rendering across 12 languages, and up to 12 consistent images per generation.
EVENTS
2026 MIT AI Conference: April 14, 2026 • Cambridge, MA
SANS AI Cybersecurity Summit 2026: April 20–27, 2026 • Arlington, VA (hybrid)
AI Dev 26 x SF: April 28-29, 2026 • San Francisco, CA
YouCode Forum: April 02, 2026 • San Francisco, CA
Lovable & Folklore VC Hackathon: April 11, 2026 • Sydney, Australia
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
