Unique Research · LLMs · December 14, 2024
Google is pushing the industry with Gemini — a family of frontier generative AI models, apps, and services. What is Gemini? How do you use it well? How does it compare with OpenAI’s ChatGPT, Meta’s Llama, and Microsoft’s Copilot? This guide tracks the latest.
What is Gemini?
Gemini is Google’s next-generation generative AI model family, built by DeepMind and Google Research. Four product lines:
- Gemini Ultra
- Gemini Pro
- Gemini Flash — leaner, faster Pro variant; plus smaller Flash-8B
- Gemini Nano — on-device (Nano-1 and fuller Nano-2), offline-capable
All Gemini models are multimodal by design — text plus audio, image, video. Google says they were pre-trained and fine-tuned on multilingual text, public and proprietary audio/image/video, and codebases. That is a break from text-only LaMDA. Caveat: training on public data without owner permission raises ethical and legal questions; Google’s AI indemnity for Cloud customers has exceptions — use care, especially commercially.
Gemini app vs Gemini models
Models and the web/mobile Gemini app (formerly Bard) are separate. The app is a chatbot-style front door — like ChatGPT or Claude. On web it is live; on Android it replaced Google Assistant; on iOS, Google and Google Search apps act as clients. Android recently added an overlay over any app (long-press power or “Hey Google”) to ask about on-screen content such as YouTube videos. The app takes images, voice, text (including PDFs; Drive video soon) and can generate images; same Google account syncs chat across mobile and web.
Gemini Advanced
Gemini features also land in Gmail, Docs, and other Google apps. Full access needs Google One AI Premium (~$20) — Gemini in Workspace (Docs, Maps, Slides, Sheets, Drive, Meet) plus Gemini Advanced (higher-tier models in the app). Perks: early features, run/edit Python in Gemini, larger context (~750k words / ~1,500 pages vs ~24k / 48 pages on free). Advanced unlocks Deep Research (plan → approve → multi-minute web research → detailed report), memory across old chats, and trip planning that uses Gmail flights, dining prefs, Maps/Search attractions and distances — auto-updating itineraries.
Enterprise: Gemini Business (Workspace add-on from ~$6/user/mo) and Gemini Enterprise (meeting notes, translated captions, doc classify/tag — custom pricing); both need annual commit.
Across Gmail, Docs, Chrome, and beyond
Gmail sidebar: draft and summarize threads. Docs: write and refine. Slides: decks and custom images. Sheets: track data, tables, formulas. Maps: café-review summaries and day itineraries. Drive: summarize files/folders. Meet: caption translation. Chrome: AI writing that respects the page context. Also: database products, cloud security, Firebase / Project IDX, Photos (natural-language search), YouTube (video ideas), NotebookLM. Code Assist (ex-Duet AI for Developers) and security products (threat intel, natural-language hunt for IoCs) run on Gemini.
Extensions and Gems
At Google I/O 2024, Advanced users got Gems — custom Gemini-powered bots from natural-language descriptions (e.g., “you’re my running coach…”), shareable or private — now in 150 countries, most languages; later Calendar, Tasks, Keep, YouTube Music. Gemini Extensions already reach Drive, Gmail, YouTube (“summarize my last three emails”); later Calendar, Keep, Tasks, YouTube Music, and Android Utilities (timers, alarms, media, flashlight, volume, Wi-Fi, Bluetooth).
Gemini Live and Imagen 3
Gemini Live: deep voice chat on the mobile app and Pixel Buds Pro 2 — even on lock screen. Interrupt mid-reply; it adapts in real time. Vision via phone camera coming. Pitched as rehearsal/brainstorm coach (interview skills, public speaking). Early reviews: promising but not yet “super useful.”
Imagen 3: built-in image gen — better prompt understanding, more detail, fewer artifacts, stronger text rendering vs Imagen 2 (per Google). People-image gen paused Feb after historical-accuracy complaints; partially restored Aug for English paid-plan pilot users.
Teens, smart home, model capabilities
June: Gemini for teens via Workspace for Education — extra policies, onboarding, AI literacy; Double Check verifies answers on the web. Gemini also lands on Google TV Streamer, Pixel 9/9 Pro, Nest Learning Thermostat; Nest Aware subscribers later get AI scene descriptions, natural-language video search, recommended automations. Nest Assistant upgrades for more natural follow-ups.
Multimodal tasks: speech transcription, live image/video captions, and more — some already shipped. Trust Google’s demos carefully (Bard launch miss; a Gemini video was not a true real-time demo). Bias and hallucination remain industry-wide.
Ultra, Pro, Flash, Nano
Ultra: multimodal homework help, paper-grounded science tasks, chart updates from formulas; native image output (not yet in all productized paths). API via Vertex AI and AI Studio.
Pro / 1.5 Pro: stronger reasoning than LaMDA; 1.5 Pro powers Advanced and can handle ~1.4M words, 2 hours video, or 22 hours audio. GA June on Vertex/AI Studio with code execution. Fine-tune or ground on Moody’s, Refinitiv, ZoomInfo, MSCI, company data, or Search; call external APIs. AI Studio templates for structured chat; Vertex Agent Builder for Gemini agents.
Flash / 2.0 Flash: once for lighter loads; 2.0 Flash is now Google’s flagship “agent-era” model — native image/audio gen, tool use (Search, APIs), faster, sometimes beating larger 1.5 models on coding/image benchmarks. Experimental on web/dev platforms; production targeted for January. Context caching available (extra fee) for large knowledge bases.
Nano: on-device on Pixel 8/8 Pro/9/9 Pro, Galaxy S24, etc. — Recorder summaries offline/private; Gboard Smart Reply; Messages Magic Compose; future call-scam warnings; weather; TalkBack object audio descriptions.
API pricing (as of Sept 2024 baseline, excl. caching)
- 1.0 Pro: $0.50 / $1.50 per 1M input/output tokens
- 1.5 Pro: ≤128K: $1.25 / $5; >128K: $2.50 / $10
- 1.5 Flash: ≤128K: $0.075 / $0.30; >128K: $0.15 / $0.60
- 1.5 Flash-8B: ≤128K: $0.0375 / $0.15; >128K: $0.075 / $0.30
~1M tokens ≈ 700k words. Ultra and 2.0 Flash pricing TBD; Nano early access.
Project Astra and iPhone?
Project Astra (DeepMind): real-time multimodal agents — Dec 2024 trusted-tester app and AR glasses prototype; not a product yet. iPhone: Apple has said it is talking with Gemini and other third parties for Apple Intelligence features; Craig Federighi confirmed talks after WWDC 2024 without detail.
— END —