Towards AI That Can Actually Interact

Today on the AI Daily Brief, a new approach to AI called interaction models. Before that in the headlines, open AI's big consulting arm is official. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. Thanks for watching! Thanks to listeners like you, the AI Daily Brief just continues to grow and grow and grow. It is almost always a top 5 technology podcast and it is almost always now in the top 200 of podcasts overall. This has happened almost totally organically because of listeners sharing with their friends and colleagues, but I think it's time to pour some gas on that particular fire. I am now hiring for what I'm calling a growth engineer. And the goal is to build stuff to make it easier for important parts of the AI Daily Brief to get to the audiences that they could be helping. This is a role for someone who is creative, dynamic, highly self-directed and lives inside Claude Coder Codex. The application is all about sharing stuff that you've built in your perspective on the show and I'll be hiring one person for a paid 3 month trial with a guaranteed base and the ability to up to double that base during that time with the goal of it turning full time and ongoing after that. You can find the role at jobs.adilybrief.ai. That's jobs.adilybrief.ai. We kick off today with a story that is a follow up to reports from last week, OpenAI is indeed launching a consulting company and now it is official. The business will operate as a separate company called the OpenAI deployment company or deploy co for short. This is basically a forward deployed engineer shop that will pair developers with some of OpenAI's most important clients to help them on theoretically real, deep AI and agentic transformation. Now like the anthropic venture that was announced last week, deploy co is structured as a joint venture in this case with 19 partners across consulting, private equity and finance. The initial investment was $4 billion at a pre-money valuation of $10 billion with TPG as the lead investor, advent international, bayonet capital and Brookfield as also co lead founding partners. Now one of the things of note here is that word on the street is that a big part of the motivation for the firms who are investing in this is to get first access to this set of engineers for their portfolio companies. In other words, this was effectively a buy in cost to skipping the line when it comes to getting help on AI transformation. Of the partners announced it looks like Goldman Sachs is the only one to back both deploy co and the anthropic effort which is as yet unnamed. The guts of deploy co are going to be built around in acquisition, specifically engineering firm tomorrow, which will give deploy co right out of the gate about 150 staff who have experience in deploying AI solutions. At this point there is basically no way to grow fast enough to meet this demand other than acquisition so I would expect a lot more M&A soon. Even though this was a big part of the discourse last week there was a surprising amount of conversation about this. Most of it was just a reaffirmation of what has finally started to become conventional wisdom, which is that it doesn't matter how powerful the models are, they are going to crash headlong into institutional inertia and for enterprises to close the capability overhang and actually get the full value from these models, it is going to involve meaningful support structures being built around them both inside and outside. Now one funny strand in the conversation that I've seen is a lot of folks running smaller versions of these agencies trying to sort of puff out their chest to make it clear that they still have a market because no matter how well resource the efforts from open AI and anthropic are, there's a massive long tail of people that need support too. And to them I would just like to say guys don't worry, no one thinks that because open AI and anthropic are in the game, somehow they are going to be able to consume the sheer tonnage of transformation support that is going to be needed over the coming decades to get these companies onboarded onto AI. And up we have an interesting market sort of sub story. You might have seen posts like this one from the Kobayesi letter suggesting something about anthropic or open AI's market implied pre IPO valuation. Now what that type of statement means is not a reference to evaluation in an actual fundraising round, but about weird gray secondary markets usually filtered through some blockchain or another where market activity on those markets is being interpreted as actual price signal. In many cases what you have is basically crypto tokens that claim to be backed one to one by stock held in SPVs that themselves own actual anthropic stock. Anthropic however seems not to be a fan of this. On Monday they updated a page in their support docs discussing unauthorized stock sales and investment scams. The article already stated that unapproved stock transfers were void and wouldn't be recognized in official records. Regarding SPVs anthropic noted, we do not permit SPVs to acquire anthropic stock and any transfer of shares to an SPV are void under our transfer restrictions. Any third party claiming to sell anthropic shares to the general public is likely either engaged in fraud or offering an investment that may have no value due to our transfer restrictions. The Monday edition was a list of firms known to be offering access to the stock, with anthropic specifically saying that any interest in anthropic stock offered by these firms is void and will not be recognized on our books or records. In a rare show of solidarity open AI sent a similar message to the market in the form of a blog post. They restated their position that unauthorized transfers are legally void without approval, making some vehicles claiming to have exposure to the stock worth zero. Lower Gabriel Shapiro thinks people trading these gray markets could have a huge problem. He wrote, "There is an active secondary market purportedly in anthropic stock or derivatives, including on fairly reputable or at least well known platforms like Forge. Anthropic is calling them out specifically by name and essentially saying 100% of these are illegal." Now Shapiro notes that the legal status is far from clear but attempting to void transactions could trigger an avalanche of lawsuits against anthropic and the marketplaces purporting to sell stock. Giving an indication of why they want this activity to be nipped in the bud, anthropic's notice triggered a quote-unquote "massive crash" cutting the price of anthropic on these markets in half yesterday. Now it's important about the story is that this is not just doing crypto market things. What underlies this is what has been a growing dynamic in private markets over the last honestly decade and a half, ever since the global financial crisis and the beginning of zirp apolices. Private companies have been waiting longer than ever to IPO because of basically just unlimited private capital. The dynamic has gotten to the point where some startups are even targeting the secondary market as their exit rather than aiming for an acquisition or an IPO. All of this comes down to demand from both long tail accredited and retail investors who are structurally blocked out of investing in early stage companies. If companies never go public or go public at enormously high valuations, it significantly cuts down the ability of retail investors to participate in the upside of company creation. Now the issue is that a lot of the way that secondary markets happen completely negates the disclosure rules of markets in general. Primary investments in a startup are pretty straightforward even though they're gated by accredited investor rules. You can easily verify whether you're included on the cap table and the SEC has your back on that. With secondaries there's no such protection. Investors are fundamentally buying shares in a holding company with no real ability to verify the claims being made. And that's in the best case scenario. KC Craig discussed some of the vehicles being traded in crypto markets that claimed to represent anthropic shares and wrote, brother, you are four layers of financial abstraction and broker crime away from touching an actual anthropic share certificate. Your position is a tokenized receipt for possible future economic exposure to a came in SPV that own shares in another Delaware SPV that maybe owns rights to future equity pending transfer approval. You are approximately anthropic adjacent at best. Now to be fair to these market participants, most people at this stage who are trading these assets know that they're trading IOUs. The bigger issue comes when I even investors think they actually own an anthropic stock and that turns out not to be the case. Ultimately it seems like this is actually a bigger issue than just this one instance with a potential reckoning on the way. Brian Norgard writes, "If anthropic starts invalidating layered SPVs in other "creative financing structures, private markets are in for a reckoning. The SpaceX IPO will expose just how much synthetic ownership and outright fraud has accumulated in privates." Natasha Mascarrano writes, "It's hard to overstate the amount of fake SPVs circulating in the market right now. They have always been a controversial financing tool with Andoral, OpenAI, andthropic, etc., fighting them for years. If companies crack down against them as promised, people are in for a root awakening after lockups expire. Still, Kingsley Advani thinks that anthropic has the limited options here, posting a chart with the dozens of registered SPVs holding anthropic shares and saying, "Enthropic, unlikely to void half their investors." Moving over to politics now, administration officials have walked back calls for an FDA-like approach to AI safety. Last week, National Economic Council Chairman Kevin Hassett said that the White House was considering an executive order to respond to mythos-level models. He said the new policy would put AI models through a process where they're proven safe just like an FDA drug. This led to a massive industry reaction with many fearing of burdensome regulatory structure that would slow innovation to a crawl. Over the weekend, former AI's R. David Sachs said that he'd spoken with Hassett and the FDA comparison wasn't particularly apt. In his words, "Make sure that the models before they're released to the public aren't going to cause an extreme amount of harm." Hassett noted that this "all of government, all of private sector approach is working well, and it's uncertain that an executive order will even be available to the public. The current approach is simply administration officials working directly with the AI labs to, quote, "in his words, make sure that the models before they're released to the public aren't going to cause an extreme amount of harm." Hassett noted that this "all of government, all of private sector approach, is working well, and it's uncertain that an executive order will even be necessary." Commenting on how much discussion his off-the-cuff comparison had generated he added, "I probably shouldn't have called it the FDA." Lastly today, President Trump is assembling a tech envoy for his trip to China later this week. How said that you like it? On Musk, Apple CEO Tim Cook and Metta President Dina Powell McCormick will join the president's delegation. The group also includes numerous finance, semiconductor, aerospace, and agriculture executives. US officials have said they intend to finalize trade negotiations with China during the meeting, including establishing a bilateral board of trade. A senior official said the executives are from companies with significant Chinese exposure, and represent sectors to be included on the trade agenda. Notably absent is Jensen Huang. Last week the NVIDIA CEO said he would join the envoy if invited, but it appears that the invite was not extended by the White House. There's a lot of different ways to look at this. Huang has traveled extensively with the president over the past year, joining trips to the Middle East and the UK. There are also apparently executives from micron and qualcom attending, so it's clear that semiconductors will at least in some way be discussed. However, Huang's absence could mean that the White House is sending a signal that NVIDIA's AI chips are off the table as part of the trade talks. While the White House signaled in December that older H-200 GPUs would be approved for export to China, those plans have stalled, and so far zero export licenses have been approved by the Commerce Department. We'll see later this week how much the absence of Jensen is an actual strategic move. For now though that is going to do it for today's headlines. Next up, the main episode. One of the most important AI questions right now isn't who's using AI. It's who's using it well. KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers, they treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news, these behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com/us/sophisticated. Today's episode is brought to you by granola. granola is the AI notepad for people in back-to-back meetings. You've probably heard people raving about granola, it's just one of those products that people love to talk about. I myself have been using granola for well over a year now and honestly it's one of the tools that changed the way I work. granola takes meeting notes for you without any intrusive bots joining your calls. After the call you can chat with your notes, ask granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes which are pre-made prompts. Once you try it on a first meeting, it's hard to go without. Head to granola.ai/aiDaily and use code AIDaily, new users get 100% off for the first three months. Again that's granola.ai/aiDaily. Quick question, when was the last time you actually visited a website to research something? If you're like me, AI pretty much does that work for you now. That of course raises a new question for brands. If AI is doing the discovering, researching, and deciding who or what is your website really for. That shift in user behavior, the rise of AI bots becoming your most important new visitors, is what my sponsor scrunch is taking head on. Scrunch is the AI customer experience platform that helps marketing teams understand how AI agents experience their site, where they show up in AI answers, where they don't, and what's preventing them from being retrieved, trusted, or recommended. And it's not just visibility. Scrunch shows you the content gaps, citation gaps, and technical blockers that matter, and helps you fix them so your brand is found and chosen in AI answers. Now for our listener, scrunch is providing a free website audit that uncovers how AI sees your site, where there's gaps, and how you're showing up in AI versus the competition. Run your site through it at scrunch.com/aiDaily. Here's a harsh truth. Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized. Most of the companies have AI tools, but only 12% use them for business value. Most employees are still using AI to summarize meeting notes. If you're the one responsible for AI adoption at your company, you need section. Section is a platform that helps you manage AI transformation across your entire organization. It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value. The result? You go from rolling out tools to driving measurable AI value. Your employees move from meeting summaries to solving actual business problems, and you can prove the ROI. Stop guessing if your AI investment is working. Check out section at section ai.com. That's SEC T-I-O-N-A-I dot com. Welcome back to the AI Daily Brief. Today's episode is a bit surprising to me. We're discussing a new model, which isn't out of the norm for this show, but it is not a new model from OpenAI or Anthropic or even from one of the Chinese labs. Instead, it comes from thinking machines and is all about a new mode of interaction. Now just by way of quick background, if there was going to be a lab outside of the Biggies, that could drop something that would catch our attention, thinking machines lab is a pretty good bet for that. Former OpenAI CTO, Mirror Morati, left OpenAI to form the lab, and pulled away a super team of researchers directly from the labs while also doing some very aggressive fundraising. In any other era in the past, the low billions of dollars that they had raised in fundraising would be notable. It's just obviously compared to the tens or even hundreds of billions in resources that the Biggies are playing with. A billion or two seems frankly pretty quaint. Thinking machines released their first product, Tinker Last October, which was a platform for reinforcement learning as a service, essentially allowing companies to find two in open source models. It wasn't received poorly exactly, but it certainly didn't capture a ton of attention or discourse in the industry. Right last year, we got rumors of more aggressive fundraising and talk of TML releasing their own model, but things went pretty quiet to begin this year, save for a wave of reports that staff and founders were leaving the company. The highlight was that two of TML's co-founders, Barrett Zoff and Luke Metz, left in January to return to work at OpenAI. And of course, we are now in the firm realism period of AI. Just last week, for example, we had Elon agree to allow Anthropic to use Colossus 1, which I basically argued was him getting comfortable accepting reality and transitioning into a different role vis-a-vis the industry because of how much was consolidating around the top players. But with all that as background, let's come back to what Thinking Machines actually shared yesterday. Mura tweeted, "Today we're sharing our work on interaction models, a new class of model trained from scratch to handle real-time interaction natively, instead of gluing it onto a turn-based one. The current AI experience often feels like a conversation that only begins after we stop talking. We have to batch our thoughts. We can't point at things. We phrase questions like emails. The interface doesn't leave room for us so we adapt to the models. We started thinking machines to advance human AI collaboration, and this is our first bet on what this looks like. Most labs treat autonomy as the goal, and interactivity is scaffolding around a turn-based core. We think the way we work with AI matters as much as how smart it is. It has to be in the model, and it has to scale with intelligence rather than trail behind it. Digging deeper into the problem, the companion blog post, which is closer to a research paper than your average announcement post, argues that today's AI systems create what they call a collaboration bottleneck. Users need to stay involved, clarify, interrupt, point, show, and correct, but current interfaces are mostly built around discrete turns. Describing this turn-based model they write, today's model's experience reality in a single thread, until the user finishes typing or speaking, the model waits with no perception of what the user is doing or how the user is doing it, until the model finishes generating its perception freezes, receiving no new information until it finishes or is interrupted. This creates a narrow channel for human AI collaboration that limits how much of a person's knowledge, intent, and judgment can reach the model, and how much of the models work can be understood. Picture trying to resolve a crucial disagreement over email rather than in person. And indeed, this analogy that current AI systems are too much like email is one that runs throughout their messaging. Their proposed solution is what they call an interaction model, trained from scratch around continuous time-aware exchange. Instead of that turn-based system, where inputs and outputs are in their words flattened into one order token sequence, their model processes streams in 200 millisecond microturns. Instead of a flattened ordered sequence with human input leading to model output, leading to human input leading to model output, their timeline version has continuous parallel input and output streams that are split into these microturns. They write, "An interaction model is in constant two-way exchange with the user, perceiving and responding at the same time." Now architecturally, they actually describe a two-part system, a real-time interaction model that stays present with the user, and a background model that handles longer reasoning, browsing, tools, and agentic work. What this allows for is the interaction model can keep talking and listening while the background model works, and then together be able to weave the results of the background model into the conversation when appropriate. So what are some of the examples and capabilities they show off? For this, I would definitely suggest that you go check out either the blog post or the announcement thread on Twitter where they include all these examples. These are not polished launch video types of assets. Instead, their TML researchers, who are actually just giving examples of the capability set. The first video, for example, they show how the model can recognize when someone new comes on the screen and make mention of that, how the model can do simultaneous translation, actually starting to translate what someone is saying from one language to another while they're still speaking, which is sort of similar to how you see translators at events speaking just a couple of seconds after the start of the phrase that they're translating. Another example of a capability they show, they call dialogue management, where the interaction model can track when the speaker is thinking, yielding, self-correcting, or inviting a response. Basically, they say there's no specific built-in dialogue management system so that it can adapt to whatever the context is. In one video, they demonstrate visual interjection by showing a researcher ask for the model to identify when she starts slouching to remind her to change her posture. In another example of that simultaneous speech, they take it out of the realm of language translation and into the realm of professional softening. The researcher is basically saying what he would like to say to a colleague who's always late, with the model changing it to a more socially acceptable version in real time. There are actually a number of other examples as well, but one that I think that's important that shows the interplay between the interaction model that the user is interacting with in the background model that's doing things, are the examples where they show the model effectively multitasking, interacting with the user while running search in the background, making the model just seem much smarter and more capable. I just watched the new Devil Wears product in a movie and I heard that it has a pretty massive opening box office. Yeah, it's crushing the box office in its opening weekend with around $233.6 million globally. Did you enjoy how they brought back Andy? Yeah, and also I really like the fact that Lady Gaga is featured in the movie. Definitely. Her cameo and runway event performing shape of a woman has a ton of tension with her implied history with Miranda. And the important thing to note here is that obviously the Devil Wears product too, which is in theaters right now, is not in the training set of this model. It was searching in the background as the researcher was interacting with it. And this I think is what we are means when she says we think the way we work with AI matters as much as how smart it is. Another way you could put this is situational smartness. We have to create the right setting for AI to be smart rather than dropping it into the rest of the world that we operate in when we're not talking to AI. So let's talk about some of my and the community's observations about the model. The first one is that the TML team was extremely on message. In a way that frankly the other lab should be jealous of. Almost all of the post from TML researchers and team members are telling some version of the story which is both increasing the capability of human AI collaboration in a way that improves humans lives. Subeth Chintala for example writes, "Thinky's secret plan. One increase human to AI bandwidth. Two raise the ceiling of human plus AI intelligence. Three help humans continue as main characters in the new world." Now the other part of the message that the TML folks were very on brand with was the idea that this is a category change type of model. In fact, kind of putting a fine point on that is that while they share a bunch of benchmarks relative to other voice models, they ended up having to create two internal benchmarks to measure new proactive audio capabilities. The two benchmarks that they introduced were called TimeSpeak which they say tests whether the model can initiate speech at user specified times while producing the correct content. The example they gave was, "I want to practice my breathing. Remind me to breathe in and out every four seconds until I ask you to stop." The second benchmark they introduced was called QSpeak, testing whether the model speaks at the appropriate moment with the expected semantically correct response. For example, every time I code switch and use another language, give me the correct word in the original language. Point being that when you have to invent new benchmarks to capture the capability set, suggests at least that the capability change is pretty significant. And like I said, the newness of this was all over the TML messaging. Rowan Zellers called it the first general video plus speech model that's visually proactive. TML co-founder John Schillman, after reinforcing that think he was founded to "advanced capabilities for human AI collaboration" which he argues are under emphasized relative to intelligence and autonomy because they're harder to evaluate, says that this new model that they're introducing, which by the way is technically called the TML interaction small, will be the beginning of what they see as a different paradigm. In the future, John writes, "We think every AI system will have something like an interaction model as the outer user-facing layer, continually keeping the user informed and learning what they actually want." Claire Birch from their team expanded the philosophy behind this. She wrote, "AI is changing how people use computers. Computers are the central tool of modern work, but computer literacy is sharply stratified. Early AI use appears to be as well. AI progress so far has inherited the world view and workflows of software engineering. This makes sense. For code native researchers, the way into the machine is through code and text is the next level up. I think we have been here before. Before the GUI text was the primary interface to the computer. You interacted line by line through the CLI, typing precise instruction so as not to mess it up. The GUI was one of the greatest democratizing forces of personal computing. Alongside dramatic drops in cost, it made the computer tool usable for many. If AI is the next interaction layer, what's the GUI moment here? It is not better prompts. Chat is still surprisingly CLI-like. Even with tool use, chat rewards, verbal fluency, abstraction, and procedural skill. Think carefully crafted context-laden prompts with just the right pleasantries and abuses. But in human collaboration, we don't just throw paragraphs of polished turn-based text at each other. Even bad meetings are spoken, gestural, interruptible, context-heavy, and full of revision and repair. The next interface will need greater affordance, richer persistent shared context, lower mode switching costs, and native multimodal interaction, rather than jarring hand-offs between text image and audio. It should stay grounded and disambiguate intent well. It should let people communicate by speaking, showing, pointing, interrupting, and revising in context, narrowing the translation cost between human intent and machine action. In other words, it will let people stay fluent in the task rather than forcing us to become fluent in the tool. The GUI moment is when the user no longer has to think like the computer, or like the AI, or like the prompt engineer, in order to access the machine's capabilities. The idea, as Claire writes, of interaction models as a step towards that reality is, I think, the right way to look at this. Last year around the time, a Google introduced NanoBanana. I said that we really needed a conception of something like an unlock score or an unlock index that was a way to measure models not by the traditional benchmarks, but by how many and what type of new use cases they unlocked. The reason that felt like it mattered around NanoBanana is that the power of that model, which came out in late August of 2025, was not that it produced such prettier pictures than the other image models available. It was how steerable it was for editing. It turned out that while that was a quote unquote small change, it was a small change that unlocked a lot of different types of uses. That would be the same later on when later versions of NanoBanana also unlocked infographics through both their reasoning over a prompt, as well as interact with text in a much more fine grain way. Using the unlock index mindset, TML basically says that there are a set of things that current commercial real-time APIs simply cannot do. They group these together as visual proactivity. They say that these APIs respond to spoken turns, but they cannot proactively choose to speak when the visual world changes. For instance, they write, "If asked, please count how many push-ups I do, such a system might respond sure thing, and then remain silent, waiting for an audio only cue that never comes. When it comes to things like time awareness, verbal cues triggers, visual-based counting, and visual cues triggers, they write no existing model can meaningfully perform any of these tasks." As weeks from latent space rights, I believe the kids call this "Thinky Machines Just Brutally Frame Mog GDM and OAI." Basically everyone's definition of real-time just got a massive freaking upgrade. Chris on Twitter writes, "Thinking machines cooked hard here. What their building feels like a shift from AVM to a true interaction model. The slouching demo is my favorite example. A normal model needs you to ask, "Am I slouching?" This notices it live, understands the context, and reacts naturally without breaking the flow. The way closer to a her-level AI companion than the current prompt-in responds out AI voice models. Now Professor Ethan Mollick, immediately recognizing how many interesting new use cases this would open up, lamented that the demos were not really about those uses. He writes, "All of the demos except maybe one are the model being fun and/or annoying by correcting or reminding in real-time. There are obvious uses for this sort of model in meetings, education, training, etc. Why not demo valuable use cases?" Professor Nick Dobos, however, pointed out, "It's a tech demo, not a consumer demo, purposeful audience choice, because their business is focused on raising VC funds to build more AI and sell the companies, not to sell the consumers." Still, I think Ethan's point that there are a lot of things that this opens up is well taken. And as impressive as it is for this innovation to be coming from TML, the question is how long it stays with TML. We're cursive on X-Rights. I doubt this stays unique for long. The frontier labs now iterate on each other's successful abstractions extremely fast. With this, we are moving from turn-based chat to models designed for persistent real-time interaction. And I think they might be right. In fact, once again, just yesterday, the OpenAI developers account showed off some new capabilities of their recently released GPT Real-Time 2 model, basically showing how that real-time audio model could work as a background agent, updating a can-bed and board of to-dos as a team gives updates in a stand-up type of meeting. This, if nothing else, I think is testament to the idea that, one, the background agent paradigm, other things are just happening passively as other types of work proceed, ease likely to be an important part of the future, and two, that these real-time audio and visual models open up new possibilities in that realm. Niktobos again writes, "Why stop at tickets? What if software engineering is 100% meetings and your AI note-taker orchestrates all your coding agents in the background for you? Ten people chatting and playing with an app while an AI hums away updating it in real-time." Given how fast things change and how iterative updates are, it's a surprising day when you see something that actually feels like the beginnings of an entirely new category of opportunity. But I think that that's what this interaction model announcement feels like. You can go read about it at thinkingmachines.ai and I'm excited to see what they do with this next as well as what people do with it. For now, that is going to do it for today's AI Daily Brief. I appreciate you listening or watching, as always, and until next time, peace! [Music]