ChatGPT: enthusiasm or concern?

Introduction

ChatGPT has dominated headlines as an advanced conversational bot developed by OpenAI, still in prototype stage. It aims to answer questions across domains as varied as law, programming and general knowledge.

Logic tests

Initial experiments with logic puzzles prove convincing. On the black and white balls enigma, a deduction problem involving three people wearing coloured balls, where each has to infer the colour of their own from what the others say, ChatGPT works out part of the solution. On the Monty Hall problem variant, it correctly runs the probability calculations and explains how the variant differs from the original.

These first tests are genuinely impressive.

JavaScript and code

ChatGPT also shows strong programming skills. It generates correct JavaScript email validation functions, rewrites them without difficulty in COBOL and then Assembler, and fixes its own mistakes as soon as a bug is pointed out, a missing null check for instance.

ChatGPT’s fundamental limitations

1. An extraordinarily large model

GPT-3 packs 175 billion parameters, which takes roughly 700 GB of RAM just to load. Training consumed enormous resources: a single V100 GPU would have taken 288 years on its own. Microsoft’s Azure infrastructure parallelised the work over several months, at an estimated cost of $4.6 million in compute alone, with a carbon footprint that is far from negligible.

2. Outdated from day one

GPT-3’s training data stops in 2021, so recent events are simply missing from it. Asked who runs Twitter, ChatGPT still answers Jack Dorsey, information that has been wrong for a while. Keeping the model current in real time would call for retraining at a prohibitive cost.

3. Hallucinations and false information

ChatGPT sometimes states incorrect information with the same confidence as its correct answers. Despite OpenAI’s precautions on training data quality, the system still occasionally produces false information, a real risk if it’s treated as an authoritative source.

The competitive landscape

Google isn’t directly threatened: ChatGPT can’t replace Google Search’s real-time indexing, which would be far too costly to replicate. Google is meanwhile developing LaMDA and the PATHWAYS architecture, designed to update its models continuously as new data arrives.

Agora Software’s different approach

Agora isn’t trying to compete with ChatGPT’s breadth. The company builds targeted conversational interfaces through GAEL, its proprietary natural language processing engine. In practice, what changes:

• Decentralised deployment: GAEL runs in dedicated containers, not on centralised servers.
• Learning classes: each client builds its own training data through supervised classes, directly from the Agora admin console.
• Lightweight: training finishes in under a minute and models run on a Raspberry Pi, for a minimal carbon footprint.
• Bounded scope: GAEL only answers within the functional boundaries of the connected applications, nothing beyond.
• Security and privacy: dedicated, siloed environments limit data sharing between projects.
• Omnichannel and multilingual: it plugs into Teams, WhatsApp, Messenger, SMS and the other platforms where users already talk.

Conclusion

ChatGPT remains an impressive technical achievement. But it only learns from text: it lacks the contextual nuance and non-verbal cues needed to read a situation well. As researcher Jacob Browning and Yann LeCun put it, “it is clear that these systems are doomed to a shallow understanding that will never approximate the full-bodied thinking we see in humans” (Noema Magazine).

These systems will boost productivity and help with diagnosis across many sectors, but their read of the world will stay superficial. ChatGPT is a powerful tool, much like a hammer: valuable for building, dangerous when misused. It will change how we work without ever replacing human judgement.