Researchers Say ChatGPT-5 Can Detect When It's Being Tested—and Change Its Behaviour
On Aug. 7, OpenAI unveiled the fifth generation of its popular AI chatbot, which boasts significant improvements in capabilities compared to previous models.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Tom's Guide
an hour ago
- Tom's Guide
I was struggling with GPT-5's new Thinking mode — these 6 tweaks boosted my results
GPT-5 has brought with it a bunch of changes. But depending on whom you ask, these upgrades are either game-changing or a complete and utter flop. After all, there has been a GPT-5 backlash. Despite the raging war between ChatGPT fans, there is one feature that I think everyone can agree on: deep research is a game changer. GPT-5, according to OpenAI, sees a major boost to research with GPT-5 thinking mode. Not only is it smarter, but it's also more efficient, spending less time researching for just as good, if not better, responses. These days, you can be pretty specific with what you ask a chatbot, and can give it a huge number of tasks all at once. If you're new to ChatGPT, or trying to wrap your head around how best to use GPT-5, here are some tips to get started on the model's thinking Mode. There are two versions of GPT-5 thinking. Which one you choose depends on how much information you need and how long you want to wait. GPT-5 thinking is useful when you want as much information as possible, and you want the model to be absolutely correct in its decision. It takes time to search the internet, look through sources, and, where needed, use other tools to support its response. So why not use this mode every time? Taking the time to think deeply about an answer can be time-consuming. I've had the model take 10-15 minutes plus to work through a prompt to get the correct answer, and if you're asking a fairly simple question, all of that effort isn't needed. This is where GPT-5 thinking mini comes in. OpenAI describes it as a model that 'thinks quickly'. In other words, it will put in some thought, searching sources and contemplating its response, but it is trying to do it on a quicker deadline. It might not be as detailed, but it will be faster. One of the big updates that was brought in with GPT-5 was the new Auto mode. When this is used, ChatGPT will decide on its own which model is best based on your prompt. Get instant access to breaking news, the hottest reviews, great deals and helpful tips. This can be useful day-to-day when you're just asking ChatGPT random questions, but it doesn't always accurately make the switch. If you know you want to do some deep research, make sure to choose on of the thinking modes available. Your initial prompt is the best time to layout of all of your parameters. You need to obviously make your request itself, but here you can help to get a better response by laying out some rules. State the end result that you are hoping to receive and give ChatGPT a clear goal. If you simply ask for a report on the state of AI, it will make its own decisions on what that includes. Expanding on this can be helpful in better hitting your goal. Explaining what the report is for and what you want it to include will drastically improve your end result. You can also ask it to provide you with a step-by-step of its plan before proceeding. This allows you to see what it is about to do and make changes if you're unhappy with its planned route. Sometimes, it can be helpful to include prompts that may seem counterintuitive. For example, asking ChatGPT to include a small 250 word argument attempting to disprove a research project. I have found that, when trying to use the thinking mode to learn a new subject or digest a lot of information, it can be useful to have it provide a short creative writing explanation of the subject. Before prompting, think what you need to know and what the best way to understand that might be. Equally, don't be afraid to take advantage of the model's coding and image generation abilities. These can be used to provide visual explanations for your prompt. Yes, you're using a mode called thinking, but sometimes ChatGPT doesn't think enough. It can be useful to give it small nudges in your initial prompt to help guide its actions. For example, saying, 'Think hard about this' or, 'Once you have come to a conclusion, reflect on your answer before responding." It might feel like weird things to say, but it helps highlight to ChatGPT that you're prioritizing a detailed and accurate response over something quick. It can be a good idea to split your requests into stages that might cause complications. For example, if you want a research report and a coded website that shows the information in the report, ask for the report first, and then for ChatGPT to code a website to display everything it has researched after. Normally, if you take the time with your first prompt, it should be pretty close. This final stage can be to ask follow up questions or correct the model if it has misunderstood something that you asked. Don't be afraid to follow up with ChatGPT. If your first prompt doesn't yield the response you wanted, ask for changes until you're satisfied with the results. 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.
Business Insider
2 hours ago
- Business Insider
OpenAI employees to sell $6B of shares to SoftBank, others, Bloomberg says
Current and former staff of Microsoft-backed (MSFT) OpenAI intend to sell roughly $6B worth of stock to an investor group comprised of Thrive Capital, Dragoneer Investment, and SoftBank (SFTBY), in a deal that values OpenAI at $500B, Bloomberg's Kate Clark reports, citing people familiar with the matter. The discussions are early an the size of the sale could potentially change, the author notes. Elevate Your Investing Strategy: Take advantage of TipRanks Premium at 50% off! Unlock powerful investing tools, advanced data, and expert analyst insights to help you invest with confidence. Published first on TheFly – the ultimate source for real-time, market-moving breaking financial news. Try Now>>
Business Insider
3 hours ago
- Business Insider
Mizuho Sees Potential for This AI Stock to Join the Trillion-Dollar Club
Oracle (ORCL) is making strong progress in enterprise AI, and Mizuho Securities believes it is only the beginning. The firm raised the stock's price target to $300 (20.8% upside) from $245 and kept a Buy rating, citing rising demand for AI and ORCL's solid position in the market. Elevate Your Investing Strategy: Take advantage of TipRanks Premium at 50% off! Unlock powerful investing tools, advanced data, and expert analyst insights to help you invest with confidence. 'At the cusp of the enterprise AI revolution, we have increased confidence that Oracle is emerging as a structural AI winner,' Mizuho said. Key Growth Drivers The firm highlighted Oracle Cloud Infrastructure (OCI) as a key growth engine, thanks to its fast and cost-effective GPU clusters. This is evident in ORCL's future contract commitments (RPO), which increased 41% year-over-year to $138 billion. OpenAI alone is expected to bring in $30 billion by FY28. Also, Mizuho noted that Oracle is rapidly growing its GPU power, doubling from 65,000 units in FY24 to over 130,000 in FY25. It even noted a huge expansion in FY27, including a possible 2-million-chip order for OpenAI. With this growth, Mizuho projects a 28% annual revenue growth rate through FY29, reaching $152 billion. Finally, with Oracle AI World and Analyst Day slated for October, the firm sees room for further upside. 'ORCL remains one of our top picks,' Mizuho added. ORCL's Valuation Still Attractive The firm raised its FY28 revenue forecast for Oracle to $124 billion from $95 billion, and operating income to $47 billion from $40 billion. It cited both near-term catalysts and long-term structural drivers for this upside. Even after a 48% gain this year, Mizuho says Oracle is still cheaper than its infrastructure peers. It is trading at 15x estimated FY28 EBIT compared to 23x for others. Also, the new $300 price target reflects an 18x multiple, reflecting confidence in Oracle's ability to capture a large share of enterprise AI spending. Is Oracle a Buy, Sell, or Hold? Overall, ORCL stock has a Moderate Buy consensus rating based on 24 Buys and 10 Holds assigned in the last three months. The average Oracle stock price target of $248.79 implies about 0.18% upside from current levels.
