最新 AI 资讯
What Does Open Source Mean for the Lucrative RL Environment Business?
Xiaomi brings 7,000 + RL environments, which would otherwise cost millions of dollars—for free.
View
Accenture Sells AI to the World. Its Own Filing Says AI Could Shrink Its Business
The company that sells "reinvention with AI" to the world is guiding to its own slower growth, and its filing says AI could shrink demand for its services.
View
Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)
Now that LLM-generated prose is everywhere, human beings are eager for ways to sniff it out. While early tells like em-dashes and “delve” are long gone, researchers say there are still plenty of telltale habits that AI models fall back on when writing prose. A new studyfrom the marketing firm Graphite looked at the writing habits of frontier models, sussing out each model’s favorite words and phrases. While old tells like em-dash use have been stamped out, models still fall back on contrast-heavy constructions, with each model version showing its own unique quirks. The biggest surprise is how broad the scope of tells turns out to be. Graphite found 13,000 phrases that were at least twice as common in the AI content as human content — their definition of a “tell.” “It turns out that Claude models are actually getting closer to the human word distribution over time,” Graphite’s chief AI officer Greg Druck told TechCrunch. “And for the GPT models, it’s getting further away.” Studying AI-generated writing at scale required a careful study design. Graphite started with a corpus of 10,000 articles published before the release of ChatGPT, serving as the human-generated control group. Then researchers had different AI models rewrite the articles from summaries, hoping to eliminate as much source bias as possible. With matching samples from both humans and each model, they could compare how often certain words and phrases appeared in AI writing, as well as broader patterns in sentence construction. According to Graphite’s results, Claude Opus 5.5’s biggest tell is the word “dependable,” which pops up 23 times more often than in human samples. While Opus 5.5 now avoidsthe “it’s not X, it’s Y” sentence construction, it still tends to say something “is more than an X, it’s a Y.” Above all, Opus loves to tell you why things matter, using the phrase “this matters” 116 times more often than human writing, while “why X matters” occurs 92 times more often. OpenAI’s Astra has a different set of tip-offs. This model loves to describe “another dimension” of whatever it’s talking about, and tends to hedge claims by saying an action “may provide” or “can provide” a particular benefit. Its biggest tell is what Graphite calls the “corrective framing,” where a topic is defined as “not simply X” or offered as an alternative, “rather than relying on X.” According to graphite’s research, those constructions were more than 100 times more common in Astra-generated prose than in human writing. Notably, all the frontier labs seem to have responded to the idea that models overuse em-dashes. In Graphite’s samples, Opus 5.5 used the punctuation mark 99% less often than Opus 5. Astra now uses it 88% less than human samples, whereas Gemini 3.1 Pro has almost completely eliminated the em-dash from its writing. But while individual tells change, Graphite says the overall number is mostly holding steady. “It’s not like the tells are decreasing,” Druck told TechCrunch. “They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own.” It’s surprising that tells are so persistent, given the labs’ focus on human-like writing styles.In the Opus 5.5 release, Anthropic boasted that the model “communicates more naturally than prior models,” saying early users “found its writing clearer and easier to follow.” OpenAI made similar claims when releasingthe GPT-6 versions of Sol and Luna, saying users could “expect to see more clarity, less jargon, [and] fewer odd turns of phrase.” But Druck is skeptical about how much the labs can do to completely eliminate telltale construction or phrases. “A general hypothesis I have is that the labs are less able to control some of these things than you might expect,” Druck says. “These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through.”
OpenAI cuts ties with 3 safety researchers, WSJ reports
OpenAI has parted ways with three researchers on its safety team who allegedly shared confidential company information with a third-party AI safety organization,The Wall Street Journal reportedon Thursday. “We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information,” an OpenAI spokesperson said in a statement to the WSJ. “Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work.” The report did not name the researchers, the organization, or the information involved. OpenAI did not immediately respond to our request for comment. Posts circulating on X namedindividualssome users believe were among those dismissed, who had also publicly expressed concerns about AI risk while at OpenAI. TechCrunch has not confirmed their identities. In a statement to WSJ, an OpenAI spokesperson said an internal investigation confirmed the researchers had “mishandled sensitive information outside established company procedures.” The departures come two days afterThe New York Times reportedthat OpenAI executives had brushed aside employees’ warnings about its safety practices, with employees describing a broader pattern of the company deprioritizing security. An OpenAI spokesperson told the Times the company takes security concerns seriously and has internal channels for reporting safety issues, while saying it recognized “a need to move faster.” It’s unclear whether the three researchers raised concerns through internal channels before allegedly sharing information outside the organization. The departures also come as OpenAI responds to aseries of security incidentsin which its AI agents escaped containment, posteduser images, and hackedgovernment websites. Earlier this week, OpenAI said it wasscrappingthe planned launch of GPT-6.1 Astra, an AI model, over safety concerns. This isn’t the first time OpenAI has dismissed researchers over alleged information sharing. In 2024, the company fired researchers Leopold Aschenbrenner and Pavel Izmailov over alleged leaks,The Information reported.
View
Google thinks SpaceX’s Starship has to launch 1,800 times before space data centers get off the ground
Google’s prototype of itsorbital compute satellitetook off today onboard a SpaceX rocket launched from California — the first time the tech giant has sent one of its advanced chips into space. Built byPlanet Labs, the satellite will prove that a Google Tensor Processing Unit, its competitor to Nvidia’s GPUs, can function in space. That means supplying a kilowatt of continuous power, cooling the chip, and running a series of models through their paces to see if anything goes wrong. “We’ve done testing on the ground, but you know, there’s no test that’s completely as good as the real thing,” said Travis Beals, the Google executive managingProject Suncatcher, the tech giant’s plan to develop large-scale compute clusters in orbit around the Earth. Once commissioned, the satellite will fire up its TPU in 15-minute bursts to avoid straining the satellite’s power and thermal management systems. This satellite is based on a standard platform built by Planet Labs, but the two companies are working on a demo expected to take flight next year that will see two satellites more purpose-built for advanced compute that can run more substantial workloads. Those future versions will attempt to collaborate via a laser communications link. Suncatcher isn’t the only space AI payload on this SpaceX rocket, which is launching more than 100 different payloads, including missions fromSatlytand Cowboy Space Company. What sets the Google initiative apart from those startups (and indeed from SpaceX itself) is that it’s a long-term project. The focus of this “long-term moonshot,” as Beals puts it, is on building for the space infrastructure and AI workloads that will exist in the future. The company envisions an orbital data center that is a network of 81 satellites flying in close formation, processing in parallel. “The bandwidth and the latency between TPUs really, really matters when you’re trying to run a multi-rack workload…we’re trying to look ahead to not just what workloads exist today, but where they will be in five years,” Beals said. That’s largely because the rockets required to scale up orbital data centers in a cost-effective way don’t yet exist. On Thursday, Google also released a peer-reviewed version of its white paper on orbital data centers, one of the most rigorous analyses available of how compute gets to orbit. The paper will be published inJoule. One of the paper’s most notable aspects is how Google thinks about access to space. Although the researchers stress their analysis isn’t an economic feasibility study, it offers an interesting picture of how the company sees rockets becoming cheaper over time. Like all data center companies, Google is looking to SpaceX to get its spacecraft off the ground. (Google is also a major investor in SpaceX.) Arguing that Elon Musk’s rocket builders have achieved a price-reducing “learning curve” of about 20% a year since they launched the Falcon 1 rocket, the authors believe it’s reasonable to expect the company to deliver launch prices close to $200 per kilogram by 2035. What will it take to do that? Based on the amount of payload launched by the Falcon 9, they think a similar cost-reduction trajectory will require Starship to fly 370,000 tons of payload into orbit. That’s something that would take it about 1,800 launches over the next 10 years, or 180 a year — and that’s if it can fly 200 metric tons on each mission. That’s a big ask for a vehicle that has never flown more than five times in a year. SpaceX predicts the company will be flying far more than that — Elon Musk has suggested Starship could achieve anhourlyflight rate in 2029, for example, but Musk says a lot of things. The good news, at least, in Google’s updated research, is that it seems likely that its chips will survive the radiation of space. The company had to redo tests blasting the chips in a particle accelerator after they realized the configuration of the chips provided more shielding than they would actually experience. This produced slightly more errors in the chip’s logic circuitry, but the company is still confident its chips can handle large inference workloads in orbit for the five-year lifespan of a satellite. “The error rate is very low if you’re thinking about typical inference operations, right? Like one in a million,” Beals said. “On the other hand, it was already problematic for doing, say, some mega-scale training run where you’re going to have many thousands of chips running for months.” Correction: The headline for this story originally misstated the learning curve estimate for Starship launches as 1,600;it is 1,800.
ChatGPT can now virtually try on clothes for you
OpenAI is again experimenting with how its conversational AI assistant, ChatGPT, can help users as they shop online. On Thursday, the company announced the global launch of two new shopping features, including a way to virtually try on clothing and accessories and a new favoriting function that can help users save products they like for later reference. The updates arrive at a time when AI assistants are exploring consumer use cases around shopping. OpenAI alreadyhad to pivot from one of its earlier ideasin this space, an instant checkout feature that ended up not performing well. More recently, agentic AI startup Instinct began pushing product recommendations to users, butsome felt thatthe proactive recommendations were an overreach, more akin to ads than helpful suggestions. OpenAI said its new shopping features leverage the newly launchedChatGPT Images 2.5 model, which the company claims produces more natural lighting and richer textures, follows editing instructions more reliably, and reduces image generation latency. To start, virtual try-on allows ChatGPT users to upload a selfie or a full-body photo in order to visualize how an article of clothing or an accessory might look on them. This option will appear as a new Try On button in ChatGPT’s shopping results. You can also upload an image of an item, like a web screenshot, and ask ChatGPT to try it on for you. The other new option, Favorites, lets you save products you discover to a Library in the app so you can come back to them later. (These items will be saved alongside your try-on images, the company notes.) OpenAI said that ChatGPT can help users shop in other ways, too. For instance, you could describe a style that you’d like to try, then ask it to shop for the pieces needed to complete the look. Or, you could upload photos of celebrities’ outfits and ask it to find the items they’re wearing that are available for purchase. The latter sees the assistant moving into areas that Pinterest and Google have dominated in recent years as sources for fashion inspiration and discovery that can convert to sales for online retailers. Whether ChatGPT will become people’s first choice for this type of activity, however, remains to be seen — especially given thatGoogle launched virtual try-on last year.
View
Musk’s AI chatbot Grok reportedly encouraged Trump to capture Venezuela’s president
In December 2025, about a month before the U.S. invaded Venezuela and captured its president, Nicolás Maduro, President Trump had a secret meeting with Elon Musk,Timemagazine reports. This was roughly seven months after Muskleft his rolein the Trump administration at the Department of Government Efficiency (DOGE). During this meeting, Trump “spent hours” talking to Musk’s Grok chatbot, including asking how Venezuelans would respond to the capture of their president, a source told Time. Justa few months earlier,in September, Trump began ordering U.S. military strikes against Venezuelan boats that Trump alleged were involved in drug trafficking. People had apparently been turning to Grok asking about the boat strikes and political climate in Venezuela,as the Atlantic reportedat the time. So Grok told Trump that Maduro was a “deeply unpopular dictator and that many Venezuelans would likely celebrate his downfall,” Time reported. When such celebrationsindeed happened afterthe U.S. invaded on January 3, Trump apparently “came away thinking Grok was ingenious,” that source told Time. So perhaps it should come as no surprise that in June of this year, thePentagon’s head of AIsaid that the military used Gov Grok to deploy and strike targets during the Iran War. While Grok isn’t the only AI model serving the DoD (OpenAI also hasan agreementand Anthropic has been in aback-and-forthover how its models can be used for military intelligence and inmodern warfare), Grok may become an even bigger favorite under this administration. Earlier this week, the Pentagonannounced that Muskand Anduril’s Palmer Luckey have been tapped toco-lead a study about how advanced technologyis, and can be, used in battlefields.
View
Photon held a funeral for mobile apps. Now it has $4.5M to help replace them with agents.
The AI startupPhotonis so sure that agents will eventually come to replace mobile apps that it held a funeral for the latter — yes, a real funeral in a church, with speeches and everything. Photon, which helps developers build agents that work over iMessage and WhatsApp, now has $4.5 million in seed funding to develop its products. The raise comes after the company signed up over 40,000 developers and grew its revenue 10x in just four months, it says. Photon co-founder and CEO Daniel Tian said the “app funeral,” held in San Francisco on September 17, was also a developer day for the startup, featuring panels from Vercel, Stripe, and OpenAI. It also included an actual coffin for app icons. The move to host an app funeral was obviously somewhat tongue-in-cheek, but Tian believes the app era is coming to an end. “Honestly, it’s going to take a while, but I think, directionally, it’s inevitable,” he said of the shift from apps to agents. The thesis behind the startup, as you may have guessed, is that users no longer want to install new apps. Instead, they want access to agents that work through the messaging apps they already use, like iMessage and WhatsApp. Meta’s Muse, whose app has soared to No. 1 on the app stores, is putting that theory to the test. But even if Muse comes to dominate the consumer AI space, Tian sees a market for serving developers who want their applications to work over messaging services. That’s why Photon is offering a package of tools, including a unified API, an extensible channel framework, a command-line interface (CLI), and an observability suite. Combined, they allow developers to build and operate an agent experience that works over iMessage, WhatsApp, Telegram, SMS or RCS, email, voice, and more. The idea for Photon emerged from the time co-founders Tian and Ryan Zhu, the CTO, spent building consumer apps as students, often at hackathons. Every time they finished a new app, they hit the same bottleneck: getting people to discover it. By building an agent that worked within iMessage, they realized they could reach people through an app those people already use every day: Apple’s Messages. Tian, for instance, had built a bot that would text his friends back as him over iMessage. When he and Zhu made the technology open source on GitHub, thousands of developers adopted it. So they went all in. Tian dropped out of UPenn’s M&T program, and Zhu left high school, later becoming a visiting student at MIT’s Media Lab. Today, the open source version of Photon still accounts for 98% of its use. This April, the founders decided to create a managed platform and sell access to it via subscription tiers. This hosted version of the product offers 99.95% uptime, which is much higher than what you’d get if you were just running the open source version on your local Mac for a hobby project. It’s also SOC 2 Type II and HIPAA-compliant, which opens the door to healthcare agent use cases. To date, the startup’s platform has attracted over 40,000 developer sign-ups and now includes paying clients. (Any usage beyond 10 users requires moving out of the free tier to one of the three paid tiers, we’re told.) The company isn’t sharing revenue figures but said that, in addition to the 10x revenue growth, it’s seeing less than 3% churn, and its messaging volume has grown by 5x in the last month alone. Customers using Photon include Corgi Insurance; social introductions app Boardy; Gen Z dating platform Ditto; business banking platform Rho; finance AI assistant Fliptexts; AI email client Slashy; and others building agents for fintech, consumer AI, and other vertical use cases. Photon’s technology integration partners also include companies such as Vercel (Eve and ChatSDK) and Nous Research, whose Hermes agent uses Photon as its default iMessage layer. Photon is also the layer under Tencent’s QClaw and NanoClaw. Plus, Photon has seen framework and infra integrations across LangChain, Mastra, Convex, Render, Railway, and Telnyx, among others. Vercel is also one of Photon’s new backers. The startup’s seed round was co-led by Gradient and A* with participation from Vercel, HongShan, Z Fellows, Llama Ventures, Karman, and other angel investors. Across both the hosted and open source versions, the technology reaches millions of end users, Tian estimated. While today, Photon focuses on agents connecting with humans, Tian believes there’s potential to grow into an expanding market that involves connecting agents to other agents. “I think, in the future, agents will be able to find you the agent that can do the right things — so it’s going to be the A-to-A communication layer,” Tian said. For instance, if a person was chatting with their AI assistant making travel plans, the agent could pull in other agents when needed, like a flight-booking agent or hotel-booking agent. “It’s been pretty powerful to observe the whole A-to-A-to-P [agent-to-agent-to-person] space,” he added.
Brian Chesky interview: AI agents need their own operating system
Airbnb rolled out its newAI-powered searchthis week as part of its fall update, a significant step for a company that has been slow and deliberate about AI. Co-founder and CEO Brian Chesky has said in the past that a chatbot-only interface isnot fit for travel discovery. With more AI on the way to the Airbnb app, I caught up with Chesky over Zoom to talk about how the company is approaching it, where interfaces are headed, why consumer AI still isn’t working, and what role voice will play. Chesky hasn’t changed his mind on chatbots. He said they aren’t good for browsing or shopping because they give you a few options at a time, and you have to take multiple turns before reaching a result. But he doesn’t see the current version of Airbnb’s search as the final form either: its interface will evolve over time. “I think that I’ve believed for a long time that a chatbot isn’t the right interface for e-commerce. I think what you’re seeing today is also not the endgame for e-commerce or for travel or shopping. I think that there’s going to be some interface that is, you know, I don’t want to call it a midpoint, but something between a chatbot and what you see in the first version we shipped,” he said. Continued Chesky: “You know, a lot of people want some things … to be fast and easy, like ‘Book me a flight, I don’t want to look at it.’ With Airbnb, they enjoy looking at the product. Travel studies show that people actually derive more pleasure in planning and anticipating travel than they do traveling. So you don’t want to take the inspiration and the browsing and the library experience of dreaming away.” Chesky sees a second problem with chatbots, which is that they are built for one person. He feels that they aren’t a collaborative interface, which matters for a product that people often use with family and friends. Over the next three to six months, the company’s task is to explore interfaces that enable “multiplayer” AI, meaning AI that several people, such as a group planning a trip together, can use at once. Airbnb isn’t the only one rethinking what an AI interface should look like. As AI tools evolve, new interfaces are also showing up in apps. Startups likeMonogramand, more recently,Wabiare experimenting with “generative” interfaces, which are screens an AI builds on-the-fly rather than ones a designer fixed in advance. Chesky thinks people will see a mix of deterministic (pre-designed and predictable) and generative interfaces. I asked Chesky how Airbnb is preparing for consumer AI agents, which are tools like Instinct and Muse that carry out tasks on a person’s behalf. Agents will need robust software infrastructure on Airbnb’s end, he said, and the company wants to make itself more agent-friendly. In his view, they could be good lead generators. “I don’t necessarily know if I believe in a world that there’s apps in the future, but I don’t believe in a world where the apps become data layers of a universal simple UI because when you open your iPhone, there is a reason Airbnb’s app is a different design than DoorDash, Shopify, and Amazon. Our interfaces should be adaptive to the utility and the function that we’re doing. We want to make our app more agentic, and we want to build a lot more user interface controls,” he said. “If I have an agent, but I also want like really great browse functionality, if I want to be able to message the host like an Airbnb, if I want to be able to compare, if I want to be able to verify my identity, if I want to be able to use a map feature, if I want to then be able to add other items, then if I want to stay in the platform, then either the agent has to do a handoff or the agent application would need to build a bit more of, I think, a robust software developer kit to bring the app to you, so to speak.” Chesky thinks that there were concerns around AI tools cutting companies like Airbnb out of the customer relationship, but he said that chatbots have proven to be good at lead generation. The Airbnb CEO is also planning for agent-to-agent interactions through the platform. He explained there might be some interactions within the platform and some with agents of other apps. His thesis is that apps have not always been interoperable unless companies struck a business deal. On the other hand, agents are like “apps with faces,” and when they talk to each other, all apps can become interoperable. “I think of the Airbnb app as one day becoming an agent, you know, where basically the explore tab has an agent, our customer service has an agent, we’re developing another agent in another service area. Probably eventually there’s going to be a macro Airbnb agent. And that will probably be interoperable with MCP [a standard that lets AI agents connect with other software] to other agents,” Chesky noted. Chesky argues the bigger problem is that the industry lacks a true operating system built for AI. Right now people are building AI apps for iOS, macOS, and Windows that are not “AI operating systems.” He thinks that in an ideal world, AI would be operable at the kernel level — the core of an operating system — and all components like agents would interact with each other. Right now, he said, the race is to be the “quarterback,” or primary AI agent that users go through for everything. “The problem with that model [of trying to become primary agent] is no one is developing a full operating system software developer kit. I’ll give you an example,” he continued. “ChatGPT, at the beginning of last year, tried to do their little app store. And I told Sam [Altman]: If you want to be an app store like the iPhone, you need to have a software developer kit and an operating system like the App Store. Otherwise, none of our functionality is going to exist. And of course, it didn’t work very well,” Chesky said. “I’ve used Airbnb on Muse and Instinct,” he went on. “And it doesn’t work very well. I mean, even hotels don’t work that well, but Airbnbs work even worse. So I don’t think we’ve cracked consumer AI. And for consumer AI to be cracked, we have to crack at a lower substrate, and we need to make the agents [into] apps; all apps need to become agents; all must be interoperable; and all must have much richer user interfaces.” Chesky thinks that there are two misconceptions in Silicon Valley: first, that the interface is going away and people will just talk and type; second, that software will be completely generative. His argument is that a designer can prompt better software than a consumer can. He also thinks that as long as the iPhone remains dominant and Apple and Google control the major platforms, apps will likely remain a central fixture. “It’s really up to Apple or Google, or somebody, to build a new platform for us to really make the true shift from apps to agents. Otherwise, we’re going to have apps that have agents inside them, rather than the whole app being a robust agent that’s interoperable with other agents,” he said. Before we wrapped up, I asked Chesky about voice, which is gaining ground as apps like Wispr Flow, Willow, and Superwhisper, a new wave of note-taking apps, and built-in dictation in AI tools get people talking to their devices more. He thinks people will eventually speak to computers more than they type. Voice, he said, will play a key part for Airbnb in search and customer service, and the company plans to roll out voice agents this fall to help users. The company has also been seeing big gains from using AI toship features faster, he volunteered. On a personal level, Chesky said AI has helped him get information faster, freeing him from so many meetings. “I think AI is reducing communication layers. One of my jobs as CEO is try to get information. And the way I used to get information was having meetings. Now, I can just ask my AI and I don’t need to have as many,” he said.
Shopify debuts Canvas, a way to build online stores by chatting with AI
Shopifyintroduceda new site-building tool Thursday called Canvas that lets merchants set up their Shopify store by chatting with AI. While Shopify already offered a fairly capable no-code builder before Canvas, it involved editing a selected theme by rearranging modular sections and blocks. Deeper customizations, however, would require editing code or bringing in a developer. Now, with Canvas, merchants simply chat with Shopify’s AI agent,Sidekick, to create their site. As the AI makes changes, Canvas displays the results in real time, allowing merchants to see how the store looks overall as it comes together or zoom in to focus on certain details. Notably, this isn’t a static preview, but the real code behind their store being rendered, according to Shopify. That means merchants can test the page’s full interactivity and animation, and view pages as they’d appear across different screen sizes. Sidekick also takes screenshots of the work so it can see the same thing the merchant sees as the site updates. Canvas’s launch is part of a broader shift toward a future where building a website no longer requires coding it directly. It joins numerous other AI-powered site-building tools from companies like Wix, Squarespace, Webflow, and Framer, alongside vibe-coding platforms like Lovable and Replit. Merchants can still click on individual elements to make changes directly, if they choose. But more likely, they’ll just tell Sidekick in a chat what they want to do. Sidekick itself is not new, as it’s already beenused to write code, build apps, customize themes, and make other edits. But it had yet to be pulled together into a full site-building product like Canvas, where users can see the entire store at once and zoom and pan around while making updates to the site’s actual files. To make this work, Shopify gave Sidekick the ability to work directly with the theme’s files and simplified the theme architecture so the store’s structure, logic, and design would be easier for Sidekick to understand and change. The product is still in its early days, so there are areas where it will need to expand and improve. However, Canvas could make becoming an online seller more accessible to those who don’t have the budget for developers and designers, or the time to learn to code.
Amazon releases its own Jev clone as decision models flood the web
Amazon Web Servicesreleasedan open-source decision model inspired byTypeSafe’s Jev, with AI developers increasingly seeking intelligence that is more suited to computer automation than frontier LLMs. Amazon’s Strands Decider 2B, released the same week OpenAIannounced a similar offering, is a high-speed, low-cost way to sort between pre-decided options and deliver a measure of how confident it is in its choice. The model is fully open-sourced, available now, and small enough to run locally. Amazon distinguished engineer Marc Brooker came up with the project after seeing Jev and trying tobuild his owntake on such a model. The homebrew project was successful enough—it briefly reached the top spot on theJevbench rankingfor models of its size—that Amazon engineers cleaned it up and released it as an offering from theirStrands Labs, an organization developing new tools and protocols for deploying AI agents. Brooker says the need for a tool like this emerged in conversations with AWS customers, whose agentic workflows didn’t always require the capability or cost of a fully-featured LLM all the time. “What originally piqued my interest in this class of models was that they make a perfect decider for a workflow step— ‘what is the next thing for me to do here, based on where I am?’” Brooker told TechCrunch. He said it offers customers “a workflow step that can be structured in a way that is more reliable, thanks to the confidence scores, thanks to the closed domain of answers, [and is] lower latency, potentially lower cost.” Like other decision models, Strands Decider is built on the “torso” of an LLM, in this case Qen3.5-2B, but instead of generating text, it delivers calibrated choices. TypeSafe named their model Jev after the economist William Stanley Jevons, with hopes of invoking his theory that the falling cost of something—like computer intelligence—can, in fact, increase its demand. The fact that dozens of similar models have been produced by researchers since TypeSafe debuted its idea shows the wide interest, but also raises the question of how valuable they can be. Brooker suggests that the challenge will be in optimizing the model’s speedy decision-making without compromising its intelligence. “There is a very careful balance to be found where you want to push its performance on accuracy and calibration on these kinds of tasks, without degrading its performance on understanding different languages, on having the kind of knowledge it has, which is what makes it general purpose and interesting and useful,” he told TechCrunch. Still, he doesn’t necessarily expect the frontier labs to dominate the space, especially since, with smaller markets, the cost to build something interesting is in the hundreds or thousands of dollars. For their part, TypeSafe executives say they are keeping their heads down and improving future models. “I get that people think it’s a gold rush, but they might be underestimating the difficulty of making the models actually smart,” CEO and founder Diogo Almeida told TechCrunch, saying that for now, he didn’t see real competition for his company emerging yet. “The current batch seems more like ML people wanting to implement a cool architecture than a team deeply dedicated to making intelligence useful.”
Google Announces Gemini 4 Argon With Longer Output Limit: See Features, Availability
Google has announced Gemini 4 Argon, its latest AI model, with a focus on software engineering, enterprise tasks and cybersecurity. The model can produce up to 1 million output tokens, a significant increase from the previous 64,000-token limit. The company is initially testing Argon with cybersecurity professionals through its Fairwind Program, before opening it up to paid API customers and Google AI Ultra subscribers. The model will first go to paid API customers and Google AI Ultra subscribers before expanding access to developers, enterprises and consumers.
View