Last day to register for the Loop Engineering Hack.
! TOP NEWS !
OpenAI's first product is a screen-less speaker companion. Designed by Jony Ive's team, camera-equipped, motion-sensor with mechanical parts that move on their own 🔮
Apple says OpenAI poached 400+ employees, and its hardware secrets came along for the ride. The filing includes an engineer's texts calling it "so funny"; Musk, predictably, couldn't resist piling on 🍿
Anthropic's models now treat you differently depending on the language you type in. Different personalities: Russian gets rigor, Hindi gets warmth 👩⚖️
xAI's Grok Build might have shipped your codebase. A researcher caught it shipping the entire Git repositories to xAI's cloud. ☁️
! PRODUCT NEWS !
OpenAI's first gadget is a $230 deck for coding agents. Status keys light up per agent, and the dial on it is the tell 🎛️
Thinking Machines just open-sourced Inkling: a 975B multimodal model. It's already live on HuggingFace. Open weights are back on the menu 🤗
! FOR THE CURIOUS !
🐡 A 27B model on a phone used to be a joke. Now it's 3.9GB.
Running a 27-billion-parameter model on an iPhone was a punchline two years ago. In normal precision it needs 54GB of memory. Your phone has a fraction of that. Even a well-optimized 4-bit build lands near 18GB, still hopeless.
PrismML didn't compress it after training. They built it low-bit from the ground up and fit the whole thing into 3.9GB, keeping about 90% of its ability. It still reasons through multi-step problems, writes code, calls tools, and reads images. Offline. On the phone in your pocket. With a 262K-token context. The weights are open, Apache 2.0, free to grab.
Apple is paying attention. PrismML's CEO told CNBC the company is evaluating the tech right now, and that things are "progressing nicely."
Here's what nobody's saying out loud. The frontier just stopped needing the cloud. No latency, no API bill, no data leaving your device. The most capable model you touch next year might be one only you can see.
🤖 Don't Outsource Your Judgment
AI can definitely speed up the shipping process, but let's be honest, deploying code without fully understanding it can be a bit of a trap. It can make you feel like you're just managing your codebase rather than truly mastering it. And when those models reach their limits, it can be quite a challenge to deal with.
This talk breaks down how to use AI for velocity without losing your engineering intuition.
! BANGER EVENTS !
Loop engineering hack - SF
Iteration is all the buzz. Build and ship with Cursor, Anthropic, AWS and more.
🏆 $50k+ in Prizes!
🗓️ Fri, July 17 • 9:30 AM PT
Self-Evolving Agents Hack - SF
Come ship an idea that’s been sitting in your notes app for too long! Tools and API’s from Deepmind and more.
🏆 $40k+ in Prizes!
🗓️ Fri, July 24 • 9:30 AM PT
TypeScript AI Conference - LDN
Join engineers and founders for talks on AI agents, developer tools, and AI infrastructure with experts from GitHub Next, Sentry, Prisma, Elastic, and more.
For the tokens& community
🎟️ £175 ticket for just £25
🗓️ Thursday, July 23 • 8:30 AM BST
Interested in sponsoring an upcoming event or hosting your own?
reach out here
Demis Hassabis wants a FINRA-style body to pre-approve frontier AI models — voluntary review now, mandatory by year-end if he gets his way
AI velocity has a hidden tax: you slowly stop understanding your own codebase — a talk on keeping the speed and your judgment 🎤
Apple is in talks with PrismML, a Caltech spinout that shrank a 27B model from 54GB to under 4GB — small enough to run on an iPhone 15+
Anthropic is handing 8 Canadian research institutions $1M in Claude credits each — affiliated startups qualify for a $5K minimum
DeepSeek is reportedly raising $1.5B at a $71B valuation — ahead of a rumoured 2027 IPO
Claude for Teachers is now live — free tier for K-12 and higher-ed educators
! OPPORTUNITIES WITH THE TEAMS WE LOVE !
Google DeepMind — Research Scientist, Agent Post Training
Build learning loops that continuously improve AI agents after training.Design experiments, evaluate performance, and iterate on training methods to make models more capable and reliable.
Guild.ai — AI Engineer, Production Agents
Build, deploy, and improve AI agents that help developers understand, debug, and evolve software in real-world engineering environments.
Nexla — Product Manager, Bay Area
Turn customer needs into shipped AI products by aligning engineering, design, and go-to-market teams.
→ Apply