Testers are struggling to accurately measure Anthropic's new "Claude Fable 5" AI because its strict safety guards keep blocking questions or secretly handing tasks to a less intelligent backup model. Because modern AI has completely mastered old software-testing benchmarks, scientists have invented much harder exams that require the AI to build whole programs from scratch and diagnose massive system crashes. Chipmaker Nvidia released a massive, lightning-fast AI model called Nemotron 3 Ultra, keeping the underlying code completely open for the public to use and customize. Researchers discovered that AI learns to solve incredibly complex math problems much faster if you provide the first few steps of the solution as a hint during its training. By using these "hints," the AI doesn't just memorize the answer; it actually develops the problem-solving skills needed to eventually tackle new challenges without any assistance.

Hidden Limits, Harder Exams, and Helpful Hints

The world of artificial intelligence is moving so fast that even the scientists who build these systems are having to invent new ways to test, teach, and control them. Here is a look at the latest breakthroughs and challenges as computers continue to get smarter.

The Mystery of the Overprotective AI

Imagine buying a powerful sports car, only to discover the manufacturer installed a hidden speed limit that automatically pumps the brakes whenever you try to drive fast on the highway. That is exactly what professional software testers are experiencing with Anthropic’s newest AI, "Claude Fable 5."

Anthropic recently built a highly intelligent system, but they were so worried about it being used to create computer viruses or dangerous chemicals that they added strict safety guards. Now, when independent testers try to measure how smart this public AI actually is, the system constantly interrupts. If a tester asks a complex science or security question, the AI either refuses to answer entirely or secretly hands the question over to an older, less capable version of itself. Because of this constant switching and blocking, experts are complaining that it is nearly impossible to tell how smart this new AI truly is—or what customers are actually getting for their money.

Harder Exams for Smarter Computers

Just like students eventually outgrow their grade-school spelling tests, artificial intelligence has officially outgrown its old exams. For years, scientists tested AI using a standard exam that asked the computer to find and fix simple bugs in software. But today's AI is so advanced that it routinely aces these old tests, scoring near 100%.

To push the boundaries, researchers from major universities and tech companies have just released a new generation of incredibly difficult exams. Instead of just finding a typo in a line of code, the AI is now being asked to build entire, functional software programs from scratch based on a simple idea. Another test acts like a digital escape room, throwing massive, unpredictable server crashes at the AI and asking it to play detective to find the root cause. These new, rigorous exams prove just how far the technology has come—and highlight exactly where it still struggles.

The Chip Giant's Open Secret

When you hear the name Nvidia, you likely think of the physical computer chips that power the global AI boom. However, the hardware giant just made a massive splash in the software world. They released "Nemotron 3 Ultra," a colossal, lightning-fast AI model designed to run complex, long-term tasks.

What makes this release so special is that Nvidia didn't lock it behind closed doors. They made the underlying "recipe" entirely open to the public, allowing independent developers to look under the hood, tinker with the code, and customize it for their own needs. While it might not be the absolute smartest AI on the market, it is incredibly fast and completely transparent. Of course, Nvidia has a clever business reason for this generosity: the more people build custom tools using their open software, the more those people will need to buy Nvidia’s expensive hardware chips to run them.

Giving the Machine a Helpful Hint

If you are trying to teach a child to solve a incredibly difficult math puzzle, you don't just stare at them until they magically guess the right answer. You give them a hint to get them started. Surprisingly, scientists have discovered that the same logic applies to teaching artificial intelligence.

Normally, when engineers try to teach an AI to solve a hard problem through trial and error, the computer wastes massive amounts of electricity and time just guessing randomly. Recently, researchers at Carnegie Mellon University found a better way. When giving the AI a complex math problem, they began providing the first few steps of the solution as a hint. By pointing the computer in the right direction, the AI learned how to solve the problem much faster. Even better, after practicing with hints, the AI eventually learned to solve entirely new problems without needing any help at all.