One Chinese company wey people no too sabi before don shake di tech world wit di way dem take train big language model and make am dey accessible, rival di big US AI companies but wit small money.

Di company, DeepSeek, wey dey for Hangzhou, talk say dem use only two months and spend less than $6 million to build dia AI model wit Nvidia H800 chips wey no too advanced. Meanwhile, OpenAI, Meta, and Google don spend billions to do di same kind AI model.

Wetin make DeepSeek stand out no be only di small money dem spend, but di open-source style wey dem use. Dis kain move don scatter di tech world wey dey grow fast-fast.

Di R1 model wey dem release under MIT licence mean say anybody fit download, change, and fine-tune am. For China, e dey follow di country content moderation rules, but outside China, people fit modify di model anyhow dem like. Perplexity AI sef don adopt dis style.

After dem announce di R1 model, di stock for American big tech companies like Nvidia and OpenAI-backed firms drop well-well, and dem lose about $1 trillion for market value.

US President Donald Trump call di move “positive development” but warn say e suppose be “wake-up call” for American industries to hold dia competitive edge.

Nvidia describe R1 as “excellent AI advancement,” while one big tech venture capitalist for Silicon Valley, Marc Andreessen, call am “AI’s Sputnik.” OpenAI CEO Sam Altman sef praise di Chinese company, say di competition dey refreshing.

Di way DeepSeek take do di AI model wit small money don make people dey wonder whether dis na di beginning of AI wey go dey accessible to everybody, no be only for big tech companies.

‘Reinforcement learning’

DeepSeek success na because dem use reinforcement learning (RL), wey no need di heavy resources wey di competitors dey use. RL dey help system learn by trial-and-error, reward di correct results, and no need pre-labelled datasets.

Di R1 paper show say di model train completely wit dis method, and e still perform like OpenAI latest models for coding, maths, and general reasoning tasks.

One research engineer for Fraunhofer IPA, Cagatay Odabasi, talk say open-source models fit dey slow to start because dem no get plenty resources, but di community contribution dey make dem strong and flexible later.

Di open-source model wey start from $0.50 per month don dey challenge di expensive, closed-source AI models, and e don become di number one free app for Apple App Store on Monday.

Odabasi talk say dis openness dey pressure AI companies to innovate, reduce price, and improve user privacy. E dey give small players and researchers chance to build on top di work and push di field forward.

Dis kain disruption dey different from wetin Sam Altman talk for 2023, when e dismiss di idea say small team wit $10 million fit compete for AI. DeepSeek don prove say innovation no dey only for big tech companies.

Odabasi add say once powerful model dey open, di community go always find new ways to use and improve am.