AI Reflections: DeepResearch, Grok and Disappointment
- By Bilal Hafeez
If you spend time on X or Twitter, you would think the release of Grok 3.0 and ChatGPT’s DeepResearch models marks the dawn of superintelligence (AGI). In reality, these advances highlight the limitations of large language models (LLMs) and how far we still are from AGI. At their core, these models are only as good as the data they are fed. They do not truly understand their responses, and the quality of your questions matters just as much as the models themselves.
Take this example: I asked, ‘What drives currencies?’ – a topic studied for decades, if not centuries. I posed the question to three cutting-edge research LLMs. Here are their abridged answers:
- ChatGPT DeepResearch: Interest rates, inflation, economic growth, trade balances, political stability, central banks, traders, capital flows.
- Grok 3.0 DeepSearch: All the above, plus public debt, terms of trade, and purchasing power parity (PPP).
- Gemini 2.0 Thinking: Similar to ChatGPT, but adds fiscal policy, unemployment, risk appetite, FX regimes, and commodity prices.
Then, I tested a custom RAG-LLM pipeline I built, incorporating data sources tailored to financial market players. Its answer included everything from the models above, plus trend-followers (bandwagon effects), investor positioning, options flows, order flow, net foreign asset positions, productivity, and savings-investment balances.
The difference is clear: better data yields better answers. Push this further with a time-sensitive question like, ‘What market factors should I watch today?’ and these models lean on news outlets like Reuters or MarketWatch. Suddenly, your ‘AGI’ is channelling journalists’ views on markets – not quite the genius you imagined.
That is not to dismiss the progress. Grok 3.0 is lightning-fast and structures answers cleanly. ChatGPT’s DeepResearch offers unmatched depth compared to its siblings. Like calculators, spreadsheets, or Python, these tools will streamline and accelerate human tasks. But human-level intelligence? We are still a long way off.