Bindu Reddy on the road to AGI: why one "smartest" model won't be enough

While Silicon Valley competes over who can most loudly promise the arrival of artificial general intelligence (AGI) by next Tuesday, Bindu Reddy suggests taking a deep breath. The Abacus.AI CEO doesn't see the industry through the rose-tinted glasses of venture capital, but through the hard reality of deploying neural networks in actual businesses. Her diagnosis is sobering: we are still at a stage where "the best model" is a title that lasts only until a competitor's next major release. Things used to be simple — there was GPT-4 and everyone else. Now the field has fragmented, and leadership in coding, creative writing, or logical reasoning keeps changing hands.

Reddy stresses that the road to AGI isn't simply a matter of adding more parameters or buying more H100 GPUs. The problem is architecture. Today's large language models remain incredibly advanced stochastic parrots. They predict the next token, but they lack what Bindu calls "agency": the ability to independently plan complex chains of actions and adjust them on the fly without human prompting. The real breakthrough toward AGI won't come when a model has read the entire internet, but when it learns to reason in real time, using internal fact-checking before it produces an answer.

The current landscape paints an interesting picture. OpenAI's GPT-4o holds on to its reputation as the most balanced product, but Anthropic's Claude 3.5 Sonnet has suddenly become the favorite among developers and anyone who values a "human" writing style and accurate code. Meanwhile, Meta's Llama 3.1 has shown that open models can breathe down the necks of the proprietary giants. Reddy believes that for a modern CTO or product lead, betting on a single model is a strategic mistake. The future belongs to orchestration, where a dedicated software layer (a router) decides which request goes to Claude and which goes to a smaller, cheaper, faster model.

Bindu draws an interesting link between AGI progress and economic viability. Training models is getting exponentially more expensive, while quality gains are starting to slow. We are approaching a wall where simply adding "more data" no longer produces a magical leap in intelligence. To get past this barrier, the industry will have to reinvent how models are trained, possibly moving away from pure supervised learning toward methods that resemble how humans learn — through trial, error, and an understanding of cause and effect. Without that, AGI will remain just a nice marketing term for raising funding rounds.

What does this mean for the rest of us? While engineers wrestle with building a digital god, we should learn to juggle what we already have. Reddy is convinced that over the next couple of years, the winners won't be the companies that build the biggest neural network, but the ones that build the best infrastructure for using this "zoo." Real intelligence isn't only about the volume of knowledge; it's also about applying the right tool at the right moment. Until models learn to do that themselves, that job stays with us.

The bottom line: AGI won't be a one-time event or a "flash." It will be a gradual transition, and right now we're stuck at a stage where models are smart but still not autonomous. Anthropic and OpenAI will keep up their arms race, but the real power is shifting toward flexibility and the ability to combine different models in a single product.

*Meta is designated an extremist organization and banned in the Russian Federation.