AI summary by TwinMind
Gemini 2.0 blew me away
Theo - t3․gg
19m 33s
Google's AI Comeback
- Surprised by Google's initial lag in AI race despite data/algorithm advantages
- Google Search won through right algorithms + data collection
- Gemini 2.0 Flash legitimately competitive, exceeded expectations
- Now enabled for free in T3 chat platform
- Flash Light version even faster with hilarious speed
Pricing Analysis Breakdown
- GPT-4: $2.50/million input, $10/million output tokens
- GPT-4 Mini: $0.15/million input, $0.60/million output (10x cheaper than GPT-4)
- GPT-o1: $15/million input, $60/million output (100x more expensive than GPT-4 Mini)
- Claude 3.5 Sonnet: $3/million input, $15/million output (caused near-bankruptcy concern)
- DeepSeek standard: $0.27/million input, $1.10/million output (cheaper than GPT-4 Mini)
- DeepSeek R1: $0.55/million input, $2.20/million output (insanely cheap for quality)
- Gemini 2.0 Flash: $0.10/million input, $0.40/million output
- Gemini 2.0 Flash Light: $0.07/million input, $0.30/million output (new cheapest model)
Performance Benchmarks
- R1 beat GPT-o1 in coding evaluations (HumanEval)
- Gemini Flash meets/beats Claude quality at 40x lower cost for output tokens
- 30x+ cost savings compared to Claude with better performance in some areas
- Google now has cheapest quality model by significant margin
- Quality comparable to reasoning models, outperforming Claude in most areas
Context Window Capabilities
- Most models: 128k tokens (~128,000 words)
- Claude Sonnet/Haiku: 200k tokens
- OpenAI o1/o3 Mini: 200k tokens
- Gemini 1.5: 2 million tokens
- Gemini 2.0: 1 million tokens
- Gemini Pro (coming): 2 million tokens
- Million token context = entire codebases fit, eliminates need for RAG systems
- At $0.10/million input tokens, million-token analysis costs only 10 cents
Reasoning Model Features
- Gemini 2.0 Flash Thinking Experimental available
- Currently free with 10/minute rate limit (experimental pricing)
- Solved Advent of Code 2021 Day 5 in 8 seconds including reasoning
- Faster than o3 Mini, much cheaper when pricing released
- Fast thinking/reasoning model with usable results
Multimodal Capabilities
- Light model: text, images, video, audio input
- Standard Flash: enhanced audio processing + built-in search
- Built-in Google Search integration (no custom tooling required)
- Other models require custom tool systems (like BrowserBase)
- Google's search advantage integrated directly into model
Technical Infrastructure
- New TPU v6 processors driving cost/speed improvements
- 3x faster inference throughput vs previous generation
- 4x improvement in training performance
- 57% more energy efficient
- Google controls: data, model development, custom chips
- Vertical integration advantage (similar to Apple's approach)
Competitive Landscape
- DeepSeek API down for 8 days (reliability concerns)
- Groq (with Q) offers 3000+ tokens/second vs OpenAI's 70 TPS
- Groq limited to Llama models, lower context windows
- Google positioned as "Apple of AI" with end-to-end control
- Only company competing on pricing, DX, quality, capabilities simultaneously
Developer Experience
- Google Cloud/Vertex still problematic UI
- AI Studio much improved interface
- API integration simple with AI SDK from Vercel
- Getting API keys from Vertex extremely difficult
- AI Studio straightforward once key obtained
- Recommendation: avoid Vertex, use Google AI Studio
