In partnership with |  |
|
|
Good morning. It’s Monday, June 22nd. |
GLM-5.2 is a new open weight model from China, and it seems to be crushing some benchmarks against the frontier models (beating GPT-5.5 on design/UI tasks, and hanging right next to Claude Opus 4.8 on long-horizon engineering) | You technically can download this model and run it locally, but you’d need serious hardware - 8x H200 GPUs (over 1TB aggregate VRAM) for decent speed, while heavily quantized 2-bit versions can squeak onto a high-end 256GB Mac Studio or a single RTX 4090. | Do high-end open weight models threaten the value of frontier labs? Remains to be seen. | -Jeff
AI Breakfast |
|
|
You read. We listen. Let us know what you think by replying to this email. |
|
The best prompt engineers aren't typing. They're talking. | | Power users figured this out early: speaking a prompt gives you 10x more context in half the time. You include the edge cases, the examples, the tone you want — because talking is fast enough that you don't skip them. | Wispr Flow captures everything you say and turns it into clean, structured text for any AI tool. Speak messy. Get polished input. Paste into ChatGPT, Claude, Cursor, or wherever you work. | 89% of messages sent with zero edits. 4x faster than typing. Works system-wide on Mac, Windows, and iPhone. | Start flowing free |
|
|
|
Testers say GPT-5.6 Pro built a 3D simulation game in a single shot |
Sam Altman went to Stanford to call out the researchers he believes held AI back. He argues that years of skepticism around scaling large language models actively slowed down the entire field. Directly challenging critics like Yann LeCun, who view LLMs as a technological dead end, Altman insists that piling on more compute, data, and parameters keeps unlocking breakthroughs in reasoning and coding. |
To prove LLMs create genuine knowledge instead of just remixing training data, he pointed to a recent OpenAI milestone where a model disproved a long-standing mathematical conjecture. While he admitted LLMs still choke on long-term planning and complex judgment, and that robotics might ultimately need world models, Anthropic CEO Dario Amodei backed him up, agreeing that raw scaling remains the primary engine for progress. |
We are already seeing what that next scaling tier looks like in practice. Developers report that GPT-5.6 Pro successfully spun up a fully functional, Sims-style 3D simulation game in a single shot. Built into a single HTML file in under 48 minutes, the project ran entirely without Codex, external coding harnesses, or multi-step agent architectures. Early testers compared the output to Anthropic’s Fable, arguing GPT-5.6 Pro holds a distinct edge in single-pass 3D design. |
Meanwhile, how we actually interact with these models is shifting from turn-based prompting to low-latency execution. Last week, we reported that OpenAI is readying a bidirectional voice upgrade with its new GPT-Bidi architecture. Now, this next-generation Bidirectional Voice Mode is rolling out in the mobile app, allowing the AI to interrupt you mid-sentence, track background speech in real time, and correct your verbal slip-ups on the fly. |
You can also watch these upgrades in action over on OpenAI Codex, which just added visual workflow recording. By letting users record onscreen actions to automate white-collar tasks, OpenAI is building a bridge between fluid multi-modal chat and autonomous desktop engineering. |
Watch: OpenAI's Codex adds visual workflow recording for white-collar task automation |
|
Frontier models and product moves |
|
Agents and the agentic stack |
|
Business, labor, and institutions |
|
Security |
|
Hardware and infrastructure |
|
Research and benchmarks |
|
Robotics |
|
Science and medicine |
|
Watch |
|
|
Grok by SpaceXAI for Word adds a sidebar AI agent to draft, research, and edit documents directly in Microsoft Word. |
Backgrind is an always-on-top, floating overlay that lets developers manage AI coding agents without breaking their workflow. |
Laguna by Poolside offers open-weight mixture-of-experts foundation models engineered specifically for long-horizon, agentic software engineering. |
oioi is an open-source cross-platform clipboard manager that summons a searchable, glassy overlay of your history via keyboard shortcut. |
Firecrawl Research Index is a daily refreshed database of arXiv papers and GitHub repos built specifically to power AI research agents. |
|
Thank you for reading today’s edition. |
|
Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter. |
Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X! |
Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months. |
