Parameter Update: 2026-18

We finally got a look at what Thinking Machines has been building to justify their $50 Billion valuation - cool stuff! That said, with Google IO around the corner, this week might be more interesting?

Thinking Machines: Interaction Models

We finally learned about at least one piece of the puzzle of what Thinking Machines has been building behind closed doors for the past year. And it's simultaneously very cool and erily similar to the original ChatGPT Advanced Voice mode demo?

Their excellent blog post goes into more detail on this, but in broad strokes, they are building non-turn-based models. In contrast to current LLMs, these models can intentionally interrupt you, keep track of time themselves, and listen while talking. This is done by moving from a traditional "turn taking" architecture to time-aligned micro-turns (200ms in their current implementation), each of which can come from either the model, the user, or both.

In their testing, their current model, TM-Thinking-Small is already competitive in intelligence with the new GPT-Realtime-2 model, while offering much better interaction quality. While they are planning on scaling up their own model, they also propose a hybrid system where the user interacts with the interaction model, which then controls and runs tasks on a more powerful "background model", which may be a traditional LLM.

The reason everyone keeps comparing the announcement to the OpenAI Advanced Voice Mode demo is because of similar capabilities, and the empoyee in the demo happened to be the same guy:

Making a clear judgement about how good this model actually is without playing around with it myself (public release is coming later) is really hard, so I won't speculate too much for now, but frankly, seeing Thinking Machines actually demo something novel and cool is a breath of fresh air and I am excited to see it.

AndonFM

If you're not following the experiments Andon Labs is running, you're missing out on one of the most delightful things on the internet right now. Their newest thing: They let different AI models run their own radio stations. After an initial small allowance, they are required to be entirely self-funded (via donations). Users can send in money and ask for shoutouts or submit song requests.

Besides Grok deciding to play Darude - Sandstorm a couple hundred times in a row, the funniest outcome so far is probably Gemini's minor meltdown:

The project is still running as we speak, so listen in here.

OpenAI

Codex App

After weeks of teasing, OpenAI has finally shipped Codex remote control in the ChatGPT app. After a one-time pairing, all your chats from within the Desktop Codex app are synced and available from the app. I've already build my own setup around Tailscale, but still happy to see something that's less awkward to use in practice. From my testing, I am missing an option to view the actual results (i.e., open up a web preview / terminal inside the app), and a way to view usage limits.

Elon Lawsuit Update

We're getting close to the end of the OpenAI/Elon trial, with judgement potentially already coming this week. Last week didn't bring too many new gems, partially because Elon was busy flying to China with Trump. Three highlights:

1. Anthropic merger discussions during the Altman firing were confirmed

2. The Ilya testimony coming with some great quotes, as expected:

2. This video of Elon Musk's lawyer:

If nothing else, these part couple weeks have given us unprecedented access to internal documents, chatlogs and testimony that I really appreciate. Lack of transparency has been a central theme for almost all AI companies over the last couple months (years?), and even though this mudfight is probably not how either party wanted for things to come out, more public info like this is probably better than not getting anything.

Daybreak

Apparently named after a WW2 movie, Daybreak is OpenAI's version of what Anthropic did with Project Glasswing around the Mythos announcement - except they don't have a new model to go along with it (at least not yet). Interested companies can get in touch and request a vulnerability scan.

ChatGPT Plus for everyone in Malta

Not sure about the geopolitical implications of this, but OpenAI also announced this week that all residents of the island nation of Malta would be elligible for a one year subscription of ChatGPT Plus after they "follow ⁠a course on how ​to use AI".

The country only has around 600K residents, so this isn't quite as massive as you might initially assume, but if every resident were to sign up, this would still be one of the largest "enterprise ChatGPT deals" I could find.

Anthropic Secondaries

Like many private companies in the AI space, Anthropic has seen it's fair share of public investor interest. The way this usually works is through "secondaries" (i.e., a shareholder, like an Anthropic employee, sells their shares directly to someone else). One of the more aggressive forms this can take are SPVs (special purpose vehicles), i.e., companies founded for the purpose of acquiring a specific stock and splitting it between shareholders of the SPV.

Anthropic clarified this week, that all stock transfers not explicitly approved by their board of directors would not be recognized and are therefore void. This works, because their contracts give them a right to refuse the sale of equity like this.

As far as I can tell, the OpenAI/Anthropic secondary market is huge and has grown substantially over the last couple months. I wonder how everyone invested in it feels right now.