In partnership with
SpaceXAI and Cursor shipped a 1.5-trillion-parameter coding model. OpenAI's voice models can now listen and talk at the same time. GPT-5.6 finally clears government review for a full Thursday launch. And Ant Group's Robbyant just taught one 6B model to control 20 different robots.
Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs with Cursor's proprietary coding data folded directly into the training mix. It ships simultaneously in Grok Build, inside Cursor itself, and via API at $2 input pricing — explicitly positioned as a cheaper, coding-first alternative to Opus-class models.
So what: This is the first frontier model built in direct partnership with a coding-tool company rather than trained in isolation and bolted onto IDEs afterward. If the Cursor-data advantage holds up in real benchmarks, it changes who has leverage in the coding-agent stack — tool companies, not just model labs.
LingBot-VLA 2.0 is Ant Group's Robbyant follow-up to LingBot-Vision (featured last issue) — a 6B-parameter vision-language-action model trained on robot trajectories spanning 20 robot configurations and 10,000 hours of egocentric footage, all folded into one unified action space. Rather than chasing a single benchmark win, the release targets three practical gaps: bridging lab-demo performance to real-world reliability, cross-embodiment generalization, and closing the loop from foundation pretraining to deployable application. Full weights and code are open-sourced.
So what: Robotics has been stuck re-training a new policy per robot body. One 6B model generalizing cleanly across 20 embodiments — and open-sourced — is the kind of unification that turns "a paper that works in one lab" into infrastructure other robotics teams can actually build on. Combined with LingBot-Vision's spatial perception, Robbyant is quietly assembling a full open-source embodied AI stack.
Own AI deployment, grow your career
Making AI actually work day to day is becoming its own job.
On July 16, hear from three people doing it: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. They'll share what the role really looks like, how it came to exist, the skills worth hiring for, and the challenges they're tackling right now. Bring your questions, since the best moments happen live.
Register to save your spot.
GPT-Live-1 and GPT-Live-1 mini are OpenAI's new real-time voice models, built to handle full-duplex conversation — talking while still listening, the way humans actually interrupt and react. It's rolling out in ChatGPT Voice starting today, using OpenAI's current frontier model for reasoning and search underneath.
So what: Every voice AI to date has been strictly turn-based under the hood, which is why it still feels like talking to a walkie-talkie. True simultaneous listen-and-speak is the missing piece for voice agents that can actually interrupt, backchannel, and feel present in a live conversation.
Netflix's TimeSeries Abstraction was hitting tail latency in the seconds on wide partitions — causing timeouts, GC pauses, and thread queueing under load. Their fix: an asynchronous, metadata-driven engine that dynamically detects and splits oversized partitions per TimeSeries ID, without a full migration. Result: read latency drops to low double-digit milliseconds.
So what: Wide partitions are a silent killer in any Cassandra-backed system at scale — including the vector/metadata stores behind a lot of production AI infra. A dynamic, per-ID repartitioning approach that doesn't require downtime is directly reusable by any team hitting the same wall.
AI Studio's Build mode can now ingest an existing GitHub repo directly and turn it into a runtime-compatible, deployable app inside the tool — no more starting every project from a blank prompt. Existing codebases become an editable, AI-assisted starting point instead of a one-way import.
So what: Most "vibe coding" tools only work well greenfield. Letting developers bring a real repo into an AI-assisted build environment is what makes these tools usable for actual existing projects, not just weekend prototypes.
AI MODELS, EXPLAINED SIMPLY
Claude vs. Gemini vs. ChatGPT vs. Qwen vs. Kimi vs. Hermes: What's Actually Different?
People say "the AI" like there's only one. That's a bit like saying "the car" — a Corolla, a Tesla, and a kit car you weld together in your garage are all "cars," but they're built by different companies, under completely different philosophies, for different reasons. Same deal with AI chatbots. Once you see the two axes that actually separate them, the whole landscape clicks into place.
The problem, first
Every week there's a new name — Claude, Gemini, ChatGPT, Qwen, Kimi, Hermes — and they all get lumped into the same bucket: "AI." But they're made by different companies, in different countries, under two very different business models. If you don't know which bucket a name belongs to, you can't reason about why it behaves the way it does, costs what it costs, or whether you're even allowed to download and run it yourself.
|
Analogy: think of it like iOS vs. Android/Linux. iOS is a polished, single-vendor product — you use it exactly as Apple ships it, and you'll never see the source code. Android/Linux is open — you can download the actual code, run it on your own hardware, and modify it. Some AI models work exactly like iOS. Others work exactly like Linux. That's the single biggest thing to know about each one. |
The two axes that explain everything
Axis 1 — Closed vs. open-weight. "Closed" means you can only use the model through the company's app or API — you'll never download it, run it offline, or peek at how it's built. "Open-weight" means the company publishes the actual model file — anyone can download it, run it on their own computer or server, fine-tune it, even remove restrictions.
Axis 2 — Who built it, and where. Three of these six come from big U.S. frontier labs. Two come from Chinese tech companies. One isn't even a foundation model at all — it's a community remix built on top of someone else's open model.
The lineup
| Name | Made by | Open or closed | Known for |
| Claude | Anthropic (US) | Closed | Careful reasoning, strong coding, agentic tool use |
| Gemini | Google (US) | Closed | Native multimodal (text/image/video/audio), deep Google integration |
| ChatGPT (GPT) | OpenAI (US) | Closed | Most widely known, huge plugin/app ecosystem |
| Qwen | Alibaba (China) | Mostly open | Strong multilingual, efficient, popular base for others to build on |
| Kimi | Moonshot AI (China) | Open | Huge context windows, competitive open "thinking" models |
| Hermes | Nous Research (US, independent) | Open | Community fine-tune of open base models, fewer content restrictions |
Version numbers change monthly in this industry — the company, the open/closed status, and the "known for" rarely do. That's why those are the three things worth memorizing, not the model number of the week.
The "aha" moment
Hermes is the one that trips people up, so it's worth calling out: it isn't built from scratch like the other five. Nous Research takes an already-open foundation model (like Meta's Llama) and re-trains it on top — closer to a custom ROM flashed onto an open Android phone than a phone built from the ground up. That's not a knock on it; it's exactly why it exists — to give people an open, less-restricted option without the billions of dollars it takes to train a frontier model from zero.
Once you can place any new name on this grid — who built it, open or closed — you already know most of what matters about it, before you've typed a single prompt.
|
Every "AI" is either a polished, closed product from a big lab (Claude, Gemini, ChatGPT) or an open file you can download and run yourself (Qwen, Kimi, Hermes). |
Quick picks, if you just want an answer
- Best all-around assistant, especially for coding: Claude
- Best for images/video and living inside Google's apps: Gemini
- Most familiar, biggest ecosystem of integrations: ChatGPT
- Want to self-host, run it privately, or pay nothing per query: Qwen or Kimi
- Want an open model with minimal content filtering: Hermes
Got a model or term you want broken down like this? Just hit reply — that's exactly what this newsletter is for.
