AI summary by TwinMind
Everything Google just announced
Theo - t3․gg
27m 46s
Gemini Model Updates and Deep Think
- Google announced Deep Think for Gemini 2.5 ecosystem - reasoning model similar to OpenAI's O1/O3
- Scored exceptionally well on benchmarks, beating O3 on code tasks
- Not released yet - Google conducting more frontier safety evaluations first
- Speed unknown, but combines 2.5 Pro speed with deep reasoning capabilities
- Gemini 2.5 Flash improvements announced
- Previously disappointing value proposition - 50-100% more expensive than 2.0 Flash for marginal improvements
- With reasoning enabled, became as expensive as O4 Mini due to high output pricing
- Google's competitive positioning slide issues
- Claimed 2.0 Flash is "24x higher intelligence per dollar than anyone in market"
- Speaker immediately clarified "than GPT-4O" then "5x higher than DeepSeek R1"
- Mixed messaging showing rushed slide preparation
- Referenced outdated March benchmark data from open platform
TPU Hardware Advances - Ironwood
- New Ironwood TPU system: 42.5 exaflops compute capacity
- World's most powerful supercomputer as of November was 1.742 exaflops
- Google's claim of 42.5 exaflops likely has caveats but still impressive
- Technical specifications:
- 7.4 terabytes per second HBM bandwidth
- 30x more power efficient than 2018 first attempt
- 2x higher performance per watt
- Competitive advantage: Only Google benefits from TPU improvements
- Nvidia GPU improvements help all AI providers
- TPU improvements only make Gemini faster and cheaper
- Enables unique capabilities other providers cannot match
Reasoning Data and Thought Summaries
- Google finally releasing reasoning/thinking data to developers
- Previously only Cursor had special API access to Gemini 2.5 Pro thinking data
- OpenAI set bad precedent by hiding reasoning data, Google followed
- OpenAI recently started sharing reasoning data, Google 1.5 months behind
- Reasoning data now available in new Gemini 2.5 Pro/Flash versions
- Enables better user experience during inference
- No more blank loading screens - users can see model thinking process
- Critical for building applications with thinking models
Additional Gemini Features
- Native audio output capability added
- Quality impressive in demos despite some connection issues
- Computer use functionality development
- Goal: Phone can look up and automate tasks on any website
- Google well-positioned to deliver this capability
- Gemini Diffusion model announced
- Performance of 2.0 Flash at 5x speed
- Extremely fast generation: ~1000 tokens per second
- Successfully translated Rust Advent of Code solutions to JavaScript/TypeScript instantly
- Uses diffusion approach: starts with noise/random text, iteratively corrects to desired output
- Instant edit feature allows rapid code translation between languages
- Early experimental access only, not as intelligent as traditional LLMs
Google Ultra Plan Pricing
- New $250/month subscription tier announced
- Presenter purchased it for testing purposes
- Includes 30TB storage (likely unused by most)
- Access to latest/highest quality Google models including Flow V3
- Standard Pro tier: 1000 credits, can only access older models like Flow V2
- Ultra tier required for newest video generation capabilities
Flow Video Generation Model
- V2 vs V3 model comparison:
- V2: Lower quality, 20 credits per generation on Pro tier
- V3: Higher quality with audio generation, requires Ultra tier
- Generation times: ~2.5-3.5 minutes per video
- Audio generation quality poor - random clicking sounds not synced to video
- Video quality competitive but not exceptional
- Better than some competitors at rendering computer screens correctly
- Still has issues with anatomical accuracy and realistic object rendering
- Scene builder feature allows linking multiple scenes into full videos
- Comparison with competitors:
- Cling V2: 5+ minutes generation time, ~$1.40 cost, CGI-like quality
- Sora: Generated images instead of video in test (user error)
Android XR and AR Glasses
- Android XR finally launching with AI-enabled glasses
- Glasses have small display area in center of lens
- Display moves with head movement (not world-locked like Vision Pro)
- Significant limitation compared to Vision Pro's spatial computing
- Demo issues during presentation due to connection problems
- Skepticism about VR/AR products based on past disappointments
- Presenter notes Vision Pro's world-locked windows as transformative experience
- Google's approach more similar to Meta Ray-Ban glasses form factor
Performance Testing and Comparisons
- Artificial Analysis cited as best site for model performance metrics
- Token per second measurements potentially misleading for reasoning models
- Models may generate extensive reasoning tokens then summarize
- Reported speed may not reflect actual user-perceived performance
- 2.0 Flash remains fastest non-reasoning model from major providers
- Gemini Diffusion's 5x speed improvement over 2.0 Flash described as "absurd"
T3 Chat Platform Promotion
- $8/month pricing vs Google's $250/month Ultra plan
- Features: access to all models, bring-your-own-key option, generous free tier
- New model switching feature during conversations
- Resilient streaming technology (maintains connection through page refreshes)
- Promo codes: "freethinking" for first 500 users (1 month free), then 50% off
- Beta version available at beta.t3.chat
