Google DeepMind CEO says AI aces math olympiad but struggles with high school problems

When smart AI feels strangely clueless

Artificial intelligence can deliver flashes of brilliance that astonish even experts, tackling once‑intractable problems demanding deep focus and creativity.

At the same time, the same systems can leave people shaking their heads in disbelief. Simple math problems from schoolbooks often confuse these advanced machines, creating a strange mix of sharp intelligence and surprising weakness.

Gold medal skills meet simple stumbles

Google’s Gemini model has proven it can stand on the world stage by winning at the International Mathematical Olympiad. Such a victory highlights how advanced these systems have become in specialized challenges that demand structured problem-solving.

Yet the celebration is quickly overshadowed when the same system stumbles on questions designed for teenagers. The mismatch creates doubt, showing that excelling in elite competitions does not guarantee consistent everyday reasoning.

A major roadblock called inconsistency

Demis Hassabis, the chief executive of Google DeepMind, often warns that inconsistency is the single greatest barrier preventing machines from matching human intelligence. Models can appear impressive one moment and completely unreliable the next.

This uneven performance creates hesitation in calling these systems truly intelligent. Until AI becomes stable across both difficult and simple tasks, the path toward reliable artificial general intelligence remains filled with uncertainty.

Why more data is not enough

Some researchers argue that providing machines with endless amounts of data and more powerful computing might push them closer to human-like intelligence. They imagine progress will naturally follow if resources keep expanding.

Hassabis disagrees with that thinking, pointing out that size is not the only measure of ability. Without improvements in reasoning and memory, scaling up alone will not fix the most glaring gaps.

Testing AI with tougher benchmarks

The current benchmarks used to test artificial intelligence often give a false sense of success. They may measure narrow areas of performance but leave many weaknesses hidden from view.

Demis Hassabis has called for a set of tougher and more challenging tests that can push these systems to their limits. Only then can researchers measure real progress and expose the areas that remain weak.

A term called jagged intelligence

Sundar Pichai, the CEO of Google, uses the phrase artificial jagged intelligence (AJI) to describe today’s uneven systems. He says machines shine brightly in certain places yet look embarrassingly flawed in others.

The term has stuck because it captures the experience people feel when using AI. It may be dazzling one moment and then oddly fragile the next, leaving users unsure how much to depend on it.

The missing pieces in reasoning

Demis Hassabis often highlights reasoning, planning, and memory as missing pieces in modern AI. These qualities form the backbone of how humans handle problems, both simple and complex.

Without them, machines may continue to give answers that look intelligent at first but fall apart under scrutiny. Strengthening these areas will be essential if AI is to move closer to true general intelligence.

Why easy flaws matter so much

Hassabis argues that advanced machines should not break down under simple questions. If a teenager can easily spot a mistake in a complex system, the credibility of that system comes into question.

Such visible flaws reveal the gap between raw performance and practical reliability. To earn trust, AI must demonstrate steady intelligence across all situations, not just in rare or carefully designed challenges.

The long road to general intelligence

Artificial general intelligence is often described as the moment when machines can think and reason as people do. Hassabis believes reaching that level will take longer than many predictions suggest.

He urges the industry to raise its standards before celebrating progress as AGI. Until systems can consistently handle everyday reasoning along with advanced problem-solving, the technology remains short of the mark.

Sam Altman questions the AGI label

OpenAI’s CEO, Sam Altman, has shifted away from calling AGI close. He once spoke about it with confidence, but today he believes the label creates more confusion than clarity.

Altman says people use different definitions of AGI, making it difficult to know when the milestone is actually reached. He prefers focusing on continuous progress rather than chasing a single definition.

Different views on intelligence goals

In the tech world, optimism and caution sit side by side. Some leaders are eager to celebrate rapid gains, while others emphasize the importance of slow, steady standards.

This difference shows how unsettled the conversation remains about intelligence. Without agreement on what counts as success, declaring victory may be premature and misleading for both the industry and the public.

A cautious timeline for the future

Demis Hassabis suggests AGI could emerge in the next five to ten years, but he also warns against rushing the declaration. The timeline depends on solving critical problems first.

Hallucinations, misinformation, and reasoning errors still appear too often. Until those gaps are closed, claiming machines have reached the level of human reasoning would not reflect the true state of progress.

Olympiad glory hides daily struggles

Winning competitions like the International Mathematical Olympiad shows that AI can master extremely difficult problems under set conditions. The achievements look spectacular on the surface.

But everyday problem-solving requires flexible thinking that these systems do not yet have. Their frequent mistakes in common scenarios remind us that impressive medals cannot hide their ongoing struggles with ordinary reasoning.

Why memory makes a difference

Human memory does more than recall information. It supports reasoning by connecting lessons from the past to new situations, allowing people to adapt quickly.

Artificial intelligence lacks this depth of memory, leaving it vulnerable to repeated mistakes. Until models gain richer memory structures, their performance will continue to feel incomplete, no matter how advanced they appear in certain tasks.

The debate is far from settled

Sergey Brin and other optimists believe AI deserves recognition for rapid progress. Hassabis and his supporters, however, argue that the bar for intelligence must be set much higher.

The split shows no sign of fading. For everyday people, the question becomes how much to trust AI in daily tasks while the experts continue their debate about definitions and milestones.

Interested in the debates shaping AI’s future? Don’t miss DeepMind’s AGI safety paper faces skepticism.

A future shaped by consistency

Every discussion about the future of AI eventually circles back to consistency. Without it, the technology feels exciting yet unreliable, dazzling at times and disappointing at others.

Solving this weakness could transform AI into a trustworthy partner for human progress. Until then, machines will remain powerful tools that still need human judgment to guide their most important decisions.

Want to know what DeepMind’s chief really thinks about Meta’s struggles? Check out what DeepMind CEO Demis Hassabis says Meta is behind in AI and scrambling to catch up.

Do you think AI will ever achieve the steady reliability needed for true intelligence? Share your thoughts in the comments.

Read More From This Brand:

• DeepMind AI Now Solves Complex Math Science Problems
• Google Search Gets an AI Makeover
• Google AI Now Understands Images Like Never Before

Don’t forget to follow us for more exclusive content right here on MSN.

If you like this story, you’ll LOVE our Free email newsletter. Join today and be the first to receive stories like these.

This slideshow was made with AI assistance and human editing.