Бинду Редди и путь к AGI: почему одной «самой умной» модели вам не хватит

While Silicon Valley competes to loudly promise the advent of artificial general intelligence (AGI) by next Tuesday, Bindu Reddy suggests taking a deep breath. The head of Abacus.AI views the industry not through the rose-tinted glasses of venture investment, but through the harsh reality of deploying neural networks in actual business. And her diagnosis sounds sobering: we are still at a stage where the "best model" is a concept that lives exactly until the next major competitor release. It used to be simple — there was GPT-4 and everyone else. Now we are witnessing fragmentation, where leadership in coding, creative writing, or logical reasoning constantly shifts from hand to hand.

Reddy emphasizes that the path to AGI is not simply a matter of increasing the number of parameters or buying new H100 graphics cards. The problem is in the architecture. Current large language models remain incredibly advanced statistical parrots. They predict the next token, but 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 to AGI will come not when a model reads the entire internet, but when it learns to reason in real time, using internal fact-checking before delivering an answer.

Looking at the current landscape, we see an interesting picture. OpenAI with their GPT-4o holds the title of the most balanced product, but Anthropic with their Claude 3.5 Sonnet have suddenly become the favorites among developers and those who value "humane" writing style and code accuracy. Meanwhile, Meta with their Llama 3.1 proved that open models can breathe down the neck of proprietary giants. Reddy believes that for a modern CTO or product leader, faith in a single model is a strategic mistake. The future lies in orchestration, where a specialized software layer (router) decides which request to send to Claude and which to a cheaper, faster smaller model.

It is interesting how Bindu connects AGI development with economic viability. Model training is becoming exponentially more expensive, and quality improvements are beginning to slow down. We are approaching a wall where simply "more data" no longer delivers a magical leap in intelligence. To overcome this barrier, the industry will have to reinvent training methods, possibly moving away from pure supervised learning toward methods that resemble how humans learn — through trial, error, and understanding of cause-and-effect relationships. Without this, AGI will remain merely a beautiful marketing term for attracting funding rounds.

What does this mean for us? While engineers struggle to create a digital god, we should learn to juggle what we have. Reddy is confident that in the next couple of years, the winners will not be those who create the largest neural network, but those who create the best infrastructure for using this "zoo" of models. True intelligence is not only a volume of knowledge, but the ability to apply the right tool at the right moment. And until models learn to do this themselves, this work remains with us.

The key takeaway: AGI will not be a one-time event or a "flash." It is a smooth transition, and right now we are stuck at a stage where models are smart but still not independent. Anthropic and OpenAI will continue the arms race, but real power is now shifting toward flexibility and the ability to combine different models in a single product.