Latest news with #TiborBlaho
Yahoo
17 hours ago
- Yahoo
GPT-5 may already be among us according to reports
When you buy through links on our articles, Future and its syndication partners may earn a commission. OpenAI appears to be testing GPT-5 based on leaked files and internal biosecurity tools GPT-5 is rumored to unify memory, reasoning, vision, and task completion into one No specific release date has been announced, but it's likely to roll out officially in the next few months Rumors that OpenAI is already testing GPT‑5 have started to spread thanks to leaks and hints appearing online. A particularly notable example was shared on X by engineer Tibor Blaho when he posted a partial screenshot of a config file hinting at 'GPT‑5 Reasoning Alpha,' dated July 13, 2025. That same week, independent researchers discovered a mention of GPT-5 in OpenAI's internal BioSec Benchmark repository, suggesting the model is already being trialed in sensitive domains like biosecurity. In case those indirect portents weren't enough, OpenAI's Xikun Zhang explicitly said that GPT-5 "is coming" during a discussion of the new ChatGPT Agent feature. The attention paid to GPT-5's release date is not surprising as people have been asking OpenAI about it since the company released GPT-4 in 2023, and only accelerated when GPT-4.5 came out last year. The ChatGPT Agents arguably brought speculation to a fever pitch, since these digital assistants can go out and do things online for you like booking tickets and organizing calendars. Those are all things people have thought GPT-5 would bring to ChatGPT. GPT-5 Future The feedback from people using ChatGPT Agents may actually be used to complete the training for GPT-5. And it will need plenty of training if the rumors about the million-token context window are true. The same goes for the supposedly unified nature of GPT-5. That would mean GPT-5 won't just include switch between features like visual analysis, code interpretation, and the same abilities as ChatGPT Agent; it will operate as a singular system. You could ask it to interpret an image, send an email, schedule a meeting, and compose and perform a vocal summary from a single prompt. For instance, a parent could coordinate school schedules, meal plans, and last-minute birthday party logistics all at once, or you could plan a trip, book hotels, put it into your calendar, and email your family the details from a single request. GPT-5 is supposedly designed to address the issues of hallucination and nuanced misunderstanding, allowing people to trust it more than GPT-4 or GPT-4.5. There's also the question of memory. OpenAI has been quietly rolling out long-term memory in ChatGPT, which means the model remembers things about you, but GPT-5 might make it even more powerful. Of course, the safety questions are already bubbling up. The mention of GPT-5 being tested in biosecurity contexts has people spooked. If it can reason about biology well enough to help with complex research, could it also spit out dangerous information if prompted the wrong way? OpenAI has promised to build in safeguards, but history tells us that people are very good at finding clever ways around digital fences. Regardless, we will all find out soon enough, though based on previous launches, it's likely to be limited to subscribers of the higher tiers of ChatGPT at first. But you can bet OpenAI won't release GPT-5 until they're sure it won't embarrass them on day one. You might also like I tried using ChatGPT Agent to plan a date night, and it worked surprisingly well OpenAI claims the new ChatGPT agent can run your errands, build your slides, and make you look like you have your life together AI took a huge leap in IQ, and now a quarter of Gen Z thinks AI is conscious
Tom's Guide
24-07-2025
- Entertainment
- Tom's Guide
ChatGPT with GPT 5 tipped for Sora 2 upgrade — here's everything we know
ChatGPT's GPT-5 upgrade is lurking right around the corner. OpenAI is hinting at it, and predictions are flooding in, both for the release day and what new features will come with it. And one new prediction could be the most exciting yet. Tibor Blaho, known for his AI predictions and insights on X, hinted at a possible update to OpenAI's Sora. He posted a screenshot of host names on which shows files with the names 'Sora 2'. Sora is OpenAI's video generator. It was originally launched back in December 2024 and, at the time, was one of the biggest jumps forward for AI video we'd seen. However, things have moved on since then, and Sora hasn't changed all that much. In fact, it has steadily been pushed to the back of the line by the likes of Google's Veo 3, as well as lesser-known competitors including Kling 2, Runway, and Moonvalley. Including an update to Sora alongside GPT-5 would seem logical. AI video is moving fast, and OpenAI won't want to get left behind. While it isn't yet confirmed, there are some features that we would expect to see from a new version of Sora. Sora 2? 21, 2025 Sora does not understand human movement. Lots of clips that you'll see from the current version depict people melting into each other, arms disappearing and bodies warping through the ground. There are a lot of reasons for this, but the big one is that AI models don't completely understand physics. They've been trained on video clips, but not the underlying explanations of those clips. Get instant access to breaking news, the hottest reviews, great deals and helpful tips. For example, footage of a man surfing makes sense in theory, but it can be hard to replicate this if you don't understand the way humans interact with water. Equally, training on video is a lot more complicated than images. Movement can be different every time you see it, and this can cause confusion for recreating the complexities of human movement. This doesn't mean it is impossible to fix. More advanced AI video models are getting close to nailing this problem, and Sora 2 could, in theory, do that too. One of the biggest factors that gave Veo 3 a lead in the market is its addition of audio with video clips. The earlier AI video generators (like Sora) were designed to just replicate video. However, more recent alternatives have begun to bring AI-generated audio into the fold as well. To make Sora 2 more competitive, OpenAI is likely to bring a feature like this into the mix. They have experimented with creating AI audio in the past but have been somewhat silent on this venture since. Where OpenAI could stand out with Sora 2 is in the length of its generated clips. Most generators are only able to produce videos between four and eight seconds at a time. However, some competitors are slowly starting to nudge this number upwards. OpenAI could be one of the companies to do this first. If they were able to do this, it would almost definitely be locked behind OpenAI's most expensive plan, considering how expensive longer video generation is and the energy costs it requires. A more recent change that is being noted in the world of AI video is an increase in the quality of the footage. Moonvalley, a company offering copyright-free AI video generation, is making waves with the quality of its footage. By training exclusively on high-quality videos, it can offer footage of only a certain resolution and above. While OpenAI is unlikely to go as far as this, improvements in both training and the technology powering the tool could see Sora 2 offering much higher quality in its video. Of course, this is tricky to blend with the longer length video mentioned above and would most likely end up being separate features.
Tom's Guide
18-07-2025
- Tom's Guide
ChatGPT is testing watermarks on AI-generated images — here's what you need to know
OpenAI has been busy recently. New features, an agent mode and now, there could be a potentially controversial update on the horizon. According to Android Authority, ChatGPT could soon be watermarking your AI-generated images. This was originally identified by X user Tibor Blaho, who spotted a mention of 'image-gen-watermark-for-free' in a beta version of ChatGPT for Android. Android Authority then followed this up with a report that the latest beta test version of the app includes a 'save without watermark' option when generating images. While it's not entirely clear what this feature will look like, or if it will come at all, it does appear that this will mean watermarked AI images, but only for those on the free version of ChatGPT. When testing the feature, Android Authority highlighted that they were able to remove the watermark using the free version of ChatGPT. However, as this is still the beta version, it is likely that this will change when the feature comes to market. With the information currently available, it is hard to know exactly what OpenAI's plan is here. It could end up being a watermark across all images, or more likely, a feature exclusively for free plans. AI providers are quickly having to find ways to make up the costs going into the models. While this used to be funding for the most part, the financial model is slowly switching to subscription plans. Get instant access to breaking news, the hottest reviews, great deals and helpful tips. This means there needs to be enough benefits for users to go up to one of ChatGPT's more expensive plans, leaving the free version behind. Having the ability to remove watermarks would be a major benefit to most users. Currently, none of the other major AI image generators are watermarking their images, so this would be a controversial move for OpenAI. Alternatively, as ChatGPT still remains the biggest competitor out there, we could simply see other players following behind. If ChatGPT does introduce watermarks on its free plan, some users will want to move elsewhere. These days, most of the best AI image generators are locked behind paywalls. The likes of Midjourney and Gemini require you to pay a subscription to create images, as does major names like Freepik. There are a few places where you can do it for free, like Grok, but these are usually only available in limited numbers. Realistically, AI image generation is slowly becoming a pay-to-use service. Or at least it will be if you want to avoid long wait times or limited numbers of image generations. Follow Tom's Guide on Google News to get our up-to-date news, how-tos, and reviews in your feeds. Make sure to click the Follow button.
Android Authority
18-07-2025
- Android Authority
ChatGPT may soon watermark your AI-generated images, but there could be a way out (APK teardown)
Calvin Wankhede / Android Authority TL;DR OpenAI could be exploring image watermarking for ChatGPT-generated images. Previous leak suggests the watermark might primarily affect free users, with a 'save without watermark' option for some. This feature is not yet official and may be subject to change before public release. OpenAI is working on a lot of new features for ChatGPT. We've spotted features like Study Together, Image Styles, and even a yearly plan in the works. Continuing the spree, we've now found code that indicates ChatGPT could be looking to watermark the images it generates, though some users could get the option to save generated images without the watermark. Authority Insights story on Android Authority. Discover You're reading anstory on Android Authority. Discover Authority Insights for more exclusive reports, app teardowns, leaks, and in-depth tech coverage you won't find anywhere else. An APK teardown helps predict features that may arrive on a service in the future based on work-in-progress code. However, it is possible that such predicted features may not make it to a public release. Back in April 2025, Tibor Blaho on X spotted that a new ImageGen watermark was in the works within ChatGPT. The app's strings mentioned the watermark as 'image-gen-watermark-for-free,' indicating that it could only be coming to ChatGPT's free users. The latest ChatGPT v1.2025.196 beta includes code that enables a 'Save without watermark' option when generating images. We managed to activate the feature to give you an early look: AssembleDebug / Android Authority As you can see in the screenshots above, ChatGPT will present the usual options for users on generated images, namely Edit, Select, Save, and Share. When the feature goes live, tapping the three-dot menu will reveal a Save without watermark option. We activated the feature on the free tier of ChatGPT, but that could be because of its in-development nature. Building upon the previous leak, it could be that the watermark is applied to all ChatGPT-generated images. However, paid subscribers could have the option to save the image without a watermark if they wish. There's also a chance that all generated images could have a watermark, and all users could have the option to save without a watermark, but the process flow would add enough friction to dissuade the casual user. ChatGPT hasn't announced this feature, so there's still a chance they could scrap the idea of watermarking images, or change other aspects of it. We'll keep you updated when we learn more. Got a tip? Talk to us! Email our staff at Email our staff at news@ . You can stay anonymous or get credit for the info, it's your choice.
Yahoo
11-04-2025
- Business
- Yahoo
OpenAI will soon phase out GPT-4 from ChatGPT
OpenAI will soon retire GPT-4, an AI model it launched over two years ago, from ChatGPT, according to a changelog posted on Thursday. Effective April 30, GPT-4 will be "fully replaced" by GPT-4o, the current default model in ChatGPT, OpenAI said. GPT-4 will remain available for use via OpenAI's API. "In head‑to‑head evaluations, [GPT-4o] consistently surpasses GPT‑4 in writing, coding, STEM, and more," wrote OpenAI in the changelog. "Recent upgrades have further improved GPT‑4o's instruction following, problem solving, and conversational flow, making it a natural successor to GPT‑4." GPT-4 was rolled out in March 2023 for ChatGPT and Microsoft's Copilot chatbot on the web. Several versions of GPT-4 had multimodal capabilities, allowing them to understand both images and text — the first for a widely deployed OpenAI model. OpenAI CEO Sam Altman has said that GPT-4, reportedly massive in size, cost more than $100 million to train. It was succeeded by GPT-4 Turbo in November 2023, a faster and cheaper model. GPT-4 is one of the models at the heart of copyright disputes between OpenAI and publishers that include The New York Times. Publishers allege that OpenAI trained GPT-4 on their data without their knowledge or consent. OpenAI claims that fair use doctrine shields it from liability. GPT-4's coming retirement will likely follow the release of new models in ChatGPT. According to reverse engineer Tibor Blaho, OpenAI is readying a family of models called GPT-4.1 — GPT-4.1-mini, GPT-4.1-nano, and GPT-4.1 — as well as the o3 "reasoning" model the company announced in December, and a new reasoning model called o4-mini. Sign in to access your portfolio
