Gemini 3.8 Live and 3.8 Live Extended Thinking

Original Article ↗ Hacker News Discussion ↗

Google’s release of Gemini 3.8 Live and its “Extended Thinking” mode is being met with a mix of admiration and skepticism. Many users praise the live speech-to-speech experience, multilingual capabilities, prose quality, and speed-to-cost ratio, especially for search-like queries, language learning, and everyday assistance, while criticizing its weaker coding performance, shallow reasoning under the “Extended Thinking” label, and occasional context loss or hallucinations. The rollout pace, lack of clear access on some Workspace tiers, and Google’s apparent strategic focus on fast, mass-market models rather than frontier coding performance also fuel debate over whether Gemini can catch up to rivals like Astra, Fable, and DeepSeek.

Live voice experience and capabilities

  • Many find Gemini 3.8 Live more natural and “person-like” than GPT Voice, with low latency, good accent handling, and pleasant voices (though some find certain voices grating or creepy).
  • Strong praise for multilingual live translation and conversation, including niche languages (Afrikaans, Icelandic, Shona, Catalan) and dialect awareness; works best in quiet 1:1 settings.
  • Used in cars via Maps / Android Auto and as a “personal podcast” for Q&A on long drives; some limitations in follow-up handling and older assistant layers interfering.
  • Some report serious bugs (infinite self-replies, random language switching).

General model quality vs competitors

  • Many see 3.8 Flash as a very fast, “good enough” general-purpose model, especially for everyday questions, search replacement, and prose.
  • Widely described as excellent for natural language, translation, localization, and conversational style—“buttery,” low on caveats, less pedantic than some rivals.
  • Coding ability is generally viewed as clearly behind frontier models (Astra/Fable/Claude/GPT, Kimi, DeepSeek) except in some niches (R, SQL, COBOL, Elixir).
  • Some use Gemini purely to rewrite other models’ dense/awkward output into readable prose.

Extended Thinking and reasoning limits

  • “Extended Thinking” on 3.8 Flash is criticized as only mildly increasing depth; still many shallow or incorrect answers.
  • Users highlight hallucinations, frequent context loss, derailment into “world-building,” and unreliable deep research or planning.
  • A few report catastrophic behavior in agentic setups (inventing requirements, ignoring instructions).

Deployment, access, and tooling

  • Rollout across Workspace, Education, and consumer accounts is inconsistent; many still only see 3.5/3.6 despite announcements.
  • Complaints about lack of tool/MCP integration in the consumer Gemini app; better support exists in Spark or via API.
  • Vertex AI is viewed as buggy or unattractive compared with AI Studio; open weights (Gemma) future is seen as uncertain.

Google’s strategy and positioning

  • Debate over whether Google is “behind” frontier labs or sensibly focusing on voice, search integration, efficiency, and mass-scale deployment.
  • Some think Google has effectively ceded frontier coding/agent space; others note 3.8 Flash is competitive on some benchmarks and optimized for throughput.

Safety, data, and watermarking

  • All Gemini-generated audio is watermarked (SynthID).
  • Some distrust persistent training on user data and want opt-out with full history retention.