Beyond language models: Why Jensen Huang dey connect open AI research with biology, weather and di future of computing
Introduction: Wetin we dey lose when dem treat AI as only chatbot?
Conversational system make artificial intelligence [AI] easy to recognise. People fit type question and read answer. But language na only one way to represent di world.
Biology, chemistry, weather and physical systems contain information wey dem no fit reduce to normal conversation. Jensen Huang argument for open AI research dey connect this wider scientific opportunity with practical computing problem: hardware designers need understand wetin future models go ask their systems to do.
Di case get several parts. Open work fit support research, increase participation and help di computing industry learn where im own architecture need change.
1. To build models dey teach computing company wetin to build
Company wey dey design AI infrastructure no fit rely only on wetin yesterday models need. E need direct experience with di work wey fit shape tomorrow systems.
Research into model architecture [Model Architecture] dey create that experience. Different approaches fit place different demands on memory, data movement, processors and di connections between dem.
Huang dey connect NVIDIA model research with im wider co-design [Co-design] strategy. To work on models provide another way to understand di systems wey those models need.
This make research valuable even when di company no dey try become di final application provider for every field. Di learning fit guide di computing platform wey others dey use.
2. Openness dey give more people room to experiment
If every useful model dey available only as finished proprietary service [Proprietary Service], researchers and smaller organisations go get less freedom to examine am, adapt am and build on top am.
Open models [Open Models] fit create another route. Researchers fit test approach, study im behaviour and investigate applications wey no be di original developer priority.
That opportunity matter across industries and countries. University laboratory, engineering team and specialist scientific group fit need different forms of access, but each one dey benefit from room to conduct im own work.
Openness no be promise say every experiment go succeed. E expand di possibility to learn and contribute without requiring everybody to begin from nothing.
3. Di word "open" need exact meaning
Model weights [Model Weights], im training code [Training Code], di data wey dem use train am and di methods wey dem use prepare that data na different things.
To make one available no automatically make di others available. Open-weight release [Open-weight Release] fit useful without providing everything wey dem need to reproduce di original training process.
For anybody wey dey choose model, di practical question na wetin di actual release permit. Di team fit inspect am, modify am, deploy am for di intended purpose and understand im limits?
Di lesson na to look beyond di label. Di useful degree of openness depend on di material wey dem provide and di conditions wey follow am.
4. Scientific AI need representations beyond normal language
Weather model dey work with di behaviour of physical system. Biological research concern structures and interactions. Robotics dey connect perception and action to environment wey dey change.
Language fit help people describe those problems, but di underlying information and relationships fit need other representations. So progress depend on more than to improve di conversational skills of large language model [Large Language Model].
Huang wider vision give specialised models place inside larger system. Agent [Agent] fit use scientific model as one of im tools, instead of trying to answer every question from language alone.
That na useful change of focus. Di aim become access to di right ability for di task, instead of expecting one model to contain every ability.
5. Platform fit support industries wey e no intend to replace
Computing company no need build di final car, discover di final drug or run every scientific institution before e fit contribute to those fields.
E fit provide models, tools and infrastructure wey help specialists do their own work. Those specialists bring domain knowledge [Domain Knowledge], experiments and practical constraints wey general platform no fit supply by itself.
This sharing of work also explain why open research and commercial products fit exist together. Proprietary application fit serve one purpose while shared research foundations support wider community.
Di boundaries suppose remain clear. Access to model na beginning, no be proof say e ready for scientific or operational decision.
Conclusion: Di next platform must make room for different kinds of intelligence
Huang case for open AI research wider than to release another chatbot. E connect computing design with scientific exploration and di ability of other people to build.
Di opportunity dey for combining general tools with specialised knowledge, while being exact about wetin open and wetin still remain di user responsibility to evaluate.
AI become more useful when im platform fit support different kinds of work. Language na important interface to that future. No be di whole future.
Powered By K7