L’Oréal has announced a collaboration with OpenAI that will bring Maybelline New York’s virtual makeup try-on feature into ChatGPT. The announcement was made at VivaTech 2026. The partnership covers consumer-facing shopping tools, product discovery, advertising pilots, research, and internal content production. The collaboration also covers L’Oréal’s internal use of AI in research, formulation, content production, and employee tools. OpenAI said in 2026 that ChatGPT had more than 900 million weekly active users and more than 50 million subscribers. Maybelline’s Makeup Virtual Try-On will be available directly within ChatGPT. The feature will use L’Oréal’s ModiFace technology, which allows users to test makeup looks digitally through a conversational interface. ModiFace is L’Oréal’s augmented reality and AI beauty technology business. L’Oréal acquired the Canadian company in 2018 to expand its digital beauty services across areas such as virtual makeup try-on, hair colour try-on, and augmented reality shopping. L’Oréal’s 2025 Annual Report said its Beauty Tech services had more than 120 million uses across 66 countries and 31 brands by the end of 2025. L’Oréal will also work with OpenAI to improve how its products are surfaced in ChatGPT in the United States. The company said the work will cover brands including Lancôme and Kérastase. L’Oréal said the ChatGPT work also includes product discovery. The company said e-commerce grew by double digits in 2025 and passed 30% of sales. Several L’Oréal brands are also involved in OpenAI’s global ChatGPT advertising pilot. They include SkinCeuticals, CeraVe, and Garnier. The programme focuses on ads within AI-assisted consumer interactions. L’Oréal described the pilot as focused on AI-native advertising at moments of consumer intent and commerce. The company has not provided further operational details on how the ad placements will appear inside ChatGPT. The partnership also extends to L’Oréal’s research work. The company said it is using GPT-Rosalind, OpenAI’s life sciences reasoning model, to map the skin microbiome. OpenAI launched GPT-Rosalind as a model for life sciences research tasks, including evidence synthesis and experimental planning. L’Oréal said it is applying the model to skin microbiome research, starting with La Roche-Posay. The skin microbiome refers to the community of microbes that live on the skin. L’Oréal said the work is aimed at identifying beneficial bacteria that can support the development of new skincare products. L’Oréal’s 2025 Annual Report also cited AI work in formulation science. L’Oréal Research & Innovation and IBM are developing a Formulation Foundation Model for beauty formulation. L’Oréal has also worked with NVIDIA on AI development and deployment. The company has said the partnership covers areas including 3D product rendering and predictive formulation science. OpenAI’s latest model will also be used in CreAItech, L’Oréal’s internal generative AI content platform. The platform is designed to create images and videos while reflecting the visual identity and history of L’Oréal’s brands. CreAItech is used by L’Oréal teams for beauty content creation. The OpenAI model support will apply to image and video generation. Asmita Dubey, L’Oréal’s chief digital and marketing officer, said the company wants to use AI to support consumers and employees. She also cited its use across marketing and research. Emmanuel Marill, OpenAI’s managing director for EMEA, said the work with L’Oréal covers research and employee tools, as well as consumer-facing services. The collaboration forms part of L’Oréal’s wider AI programme. The company said the programme covers consumer tools and internal work across marketing and research. L’Oréal said 73,000 employees have already been trained in generative AI. The company has also introduced internal tools including L’OréalGPT and personal AI companions. The announcement coincides with L’Oréal’s 10th year at VivaTech. (Photo by Helio E. López Vega) See also: Microsoft sells OpenAI models in China. OpenAI and Anthropic won’t. Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information. AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here. The post L’Oréal brings Maybelline virtual try-on to ChatGPT appeared first on AI News.
Sakana AI just pulled a neat trick with their new Fugu system. Instead of leaning on a single model, it dynamically stitches together a handful of LLMs, letting each play to its strengths as a question rolls in. The result feels like a small orchestra, each instrument stepping in when it’s the right fit, and the whole thing stays in sync without a human conductor.
What’s clever is how they’ve built the routing layer—tiny decision nets that watch the prompt, predict which model will answer best, and hand off the work in milliseconds. It’s a bit like a traffic light that knows the exact timing for every car, keeping the flow smooth and avoiding bottlenecks.
Because the system can swap providers on the fly, it reduces reliance on any one vendor. If one model hits a limit or spikes in cost, Fugu simply reroutes to another without breaking the user experience.
In early tests, the ensemble hits Anthropic’s Fable and Mythos benchmark scores, matching the numbers that previously required a single, massive model. It’s a modest but tidy proof that coordination can close the gap without scaling a single engine.
On the markets — Kalshi traders have been actively repricing this story in the last day.
Samsung Electronics is deploying ChatGPT Enterprise and Codex to all employees in South Korea and everyone in its Device eXperience (DX) division worldwide. The article Samsung rolls out ChatGPT Enterprise and Codex to employees in South Korea appeared first on The Decoder.
Bob Starr was delighted with his vibe-coded website. "Boomberg" showed how much US tax money is going to tech companies, and Starr launched it online immediately after making it. It wasn't until months after the site went live that he realized there was a problem: a hidden SQL injection risk. It could've left the site open for an attacker to read or alter data they shouldn't have access to. "It was just a glaring oversight on my part. It was a complete blindspot in my state of learning this new technology and understanding it, and I'm sure there are others making the same mistake," said Starr, a project manager in the tech sector.
I’ve been playing with NLTK’s MWETokenizer lately, and it’s a neat fix for keeping multi‑word expressions together. Instead of splitting “New York” into two tokens, you feed a list of phrases and the tokenizer merges them on the fly, which makes downstream analysis—like frequency counts or topic models—much cleaner.
The second trick is to feed POS tags into the lemmatizer. By mapping NLTK’s tag set to WordNet’s, the lemmatizer can decide whether “running” should become “run” or stay as a verb, and “better” can be reduced to “good” when it’s an adjective. It’s a small step that prevents a lot of nonsense stems.
Finally, for collocations I’ve been using NLTK’s association measures. You pull bigrams, compute PMI or Dice, and filter by a threshold. It surfaces phrases that actually co‑occur more often than chance, giving you a data‑driven way to spot idioms or domain‑specific terms without hand‑crafting lists.
So, Sakana AI has developed this system called Fugu, which is designed to help companies avoid relying too heavily on a single AI vendor. It does this by creating a pool of different models that can work together to complete tasks. The system can dynamically route traffic around any restricted or degraded provider, which helps maintain service continuity. This is particularly important given the recent export controls affecting certain AI models, which can suddenly become unavailable due to foreign policy decisions.
What's interesting is that Fugu can function as a single endpoint, but behind the scenes, it's assembling a team of expert models to work on a task. This allows companies to access top-tier computing capabilities without being locked into a single vendor. The system has been tested by almost 500 early users, and it's shown promising results in areas like cybersecurity and code review. For example, it was able to automate complete security assessment cycles and identify logic flaws and security vulnerabilities in complex codebases.
Fugu comes in two tiers: a standard model that prioritizes low latency for daily tasks, and an Ultra model that's designed for more complex, multi-step analytical problems. The Ultra model has been shown to perform competitively against leading closed models, and it's been used in areas like academic paper reproduction and patent analysis. One of the key benefits of Fugu is its ability to maintain strict compliance boundaries while executing complex tasks, which is a major advantage in areas like cybersecurity.
The system has also been used in data science units to explore mathematical hypotheses and execute experimental code runs. It's been able to sustain progress over extended periods with minimal human intervention, which is a significant advantage over single-call models that require constant human prompting. Overall, Fugu seems like a really interesting solution for companies that want to avoid vendor lock-in and maintain flexibility in their AI deployments.
So I was looking into this deal between SpaceX and Reflection AI, and what caught my attention is that Reflection AI is basically renting out a huge chunk of SpaceX's data center space, specifically the Colossus 2 near Memphis, Tennessee. They're getting access to Nvidia's latest AI chips, the GB300, which is a big deal because those chips are super powerful and can handle a lot of complex computations. What's interesting is that Reflection AI is paying a whopping $150 million per month for this access, and they've committed to doing so for the next three years, starting July 1, 2026.
This arrangement is pretty unique because it shows how companies are now looking at data centers and computing power as something you can essentially rent or lease, rather than having to build and maintain your own infrastructure. It's almost like a cloud computing service, but instead of just storing data, you're getting access to extremely powerful hardware that can handle things like AI processing.
I think what's also worth noting is that Reflection AI is an open-source AI lab, so they're not just some big corporation looking to use this computing power for their own gain. They're actually working on developing AI technologies that are open and accessible to everyone, which could have some really interesting implications for the future of AI research and development.
It's going to be interesting to see how this deal plays out and what kind of projects Reflection AI is able to work on with this kind of computing power at their disposal. The fact that they're willing to pay so much per month suggests that they have some big plans in the works, and I'm curious to see what they'll be able to accomplish.
So I was reading about how some companies are using this thing called Vibecoding to evaluate potential software acquisitions. Basically, they're using it to create AI replicas of the target company's software, which is pretty interesting. This replica can then be used to assess the competitive advantage of the target company, and it's already starting to influence purchasing decisions. It's like they're testing the software's underlying mechanics to see if it's really as valuable as it seems.
This approach is allowing companies to get a much deeper understanding of the software they're considering acquiring, and it's becoming a key factor in their decision-making process. It's not just about looking at the surface-level features and functionality, but really digging into the underlying code and architecture to see what makes it tick.
It's a pretty clever way to evaluate software, and it's giving companies a lot more confidence in their acquisition decisions. By creating these AI replicas, they can test the software's performance, scalability, and security, and get a much better sense of its overall value.
I think what's really interesting about this is that it's changing the way companies think about software acquisitions. It's no longer just about buying a product, but about buying a set of capabilities and technologies that can be integrated into their own systems. And Vibecoding is becoming a key tool in that process.
So I was reading about how the Five Eyes intelligence alliance is warning that these really advanced AI models are going to be available in just a few months. What's interesting is that these models have the potential to completely change the way offensive cyber operations are carried out. Essentially, they could be used to take down entire governments or businesses, which is a pretty significant concern.
The thing that's mechanically surprising here is that these AI models are getting to the point where they can autonomously find and exploit vulnerabilities in systems. This means that instead of needing a team of skilled hackers to carry out a cyber attack, an AI model could potentially do the same thing on its own. This changes the landscape of cyber warfare and makes it a lot more accessible to a wider range of actors.
It's not just the fact that these models exist that's the problem, it's also the fact that they could be used by anyone with the right resources. This includes nation-states, but also potentially other groups like terrorist organizations or even individual hackers. The Five Eyes alliance is made up of intelligence agencies from the US, UK, Canada, Australia, and New Zealand, and they're all warning about the potential risks of these AI models.
What's also worth noting is that these models are not just limited to carrying out cyber attacks. They could also be used to defend against them, which is an interesting dynamic. On the one hand, you have the potential for these models to be used to carry out really devastating attacks. On the other hand, you also have the potential for them to be used to make systems more secure and resilient. It's a complex issue, and it's going to be interesting to see how it all plays out.
I was surprised to hear that DeepMind isn’t just crunching numbers anymore—it’s actually sitting down with A24, the indie studio behind “Moonlight” and “Her.” Instead of a one‑off demo, they’ve signed a long‑term research pact, so the two teams can tinker with how AI might help script drafts, shot planning, or even the subtle rhythm of editing. What’s more, Google is backing the partnership with about $75 million, which feels less like a cash grab and more like a vote of confidence that the tech could genuinely smooth the creative workflow.
The idea is to keep the tools low‑key, letting filmmakers test ideas without the hype of a full‑blown AI director. DeepMind will bring its heavy‑weight models, while A24 supplies the on‑the‑ground storytelling instincts. It’s a kind of sandbox where the AI learns from real artistic choices, not just synthetic data.
If it works, we might start seeing drafts that already respect a director’s visual language, or editing suggestions that feel like a trusted collaborator rather than a robot. It’s still early, but the fact that a studio known for daring stories is willing to experiment with AI says a lot about where the conversation is heading.
Send this story to anyone — or drop the embed into a blog post, Substack, Notion page. Every play sends rev-share back to storyflo · A.I..
We’ve simplified responses to 👍 / 👎. Past comments are archived but no longer visible.
