I Spent a Week with Gemini Pro 1.5–It's Fantastic

Model capabilities & context window

• Discussion corrects the article: GPT‑4 Turbo has 128k context via API, 16–32k in web/app; Gemini 1.5’s 1M–10M context is still seen as a genuine step up.
• Some speculate Google uses brute-force scaling plus systems tricks; others mention techniques like LongRoPE, Mamba, FlashAttention, PagedAttention, but no one knows for sure.
• Several note Gemini 1.5 reportedly uses long context better than current models, not just accepts it.

Use cases for huge context

• Strong enthusiasm for feeding entire books, codebases, and large document sets:
  • E‑reader helpers (character lookup, spoiler‑free recaps, acronym reminders, motive analysis).
  • In‑depth study aids and “group of advisors” syntheses from many textbooks/advice books.
  • Codebase-wide feature suggestions and more context‑aware coding tools.
  • Large‑scale research synthesis or meta‑analysis across many papers.
• Some point out Kindle X‑Ray and similar features as partial precedents.

Cost, access & scalability

• Concern that long-context calls are extremely expensive today; people doing chat‑summary or Telegram‑summary workflows find GPT‑4 costs balloon quickly.
• Mixed views on future costs: some expect near‑zero prices over time (analogy to mainframes), others counter that Gemini may be more expensive than alternatives and worry about selective access.
• Performance under full Google‑scale load is uncertain; even the article notes current latency is high.

Hallucinations, accuracy & trust

• Multiple examples of severe hallucinations (fabricated anecdote from a biography; invented donors in a campaign-finance PDF; bogus methods in code).
• People worry that even sophisticated users are publishing unverified AI content and that readers increasingly treat LLM output as authoritative.
• Consensus: models can be powerful, but anything factual or high‑stakes must be checked, ideally by human experts.

Bias, guardrails & “wokeness”

• Intense debate over Gemini’s earlier image‑generation failures and broader “DEI/guardrails”:
  • Some see systematic, ideologically driven bias and worry it affects search and text outputs.
  • Others argue most of the problem is an over‑aggressive system prompt rather than purged training data.
  • A minority explicitly like strong inclusivity, preferring it to the opposite; others refuse to use “openly racist” or politically skewed systems.
• Skepticism that “pro” or private versions will be meaningfully different in this respect.

RAG vs long context

• Some claim long context will reduce the need for chunking-heavy RAG; you can just stuff full documents in.
• Others argue RAG remains crucial:
  • Main bottleneck is intelligent selection of relevant material, not raw window size.
  • For ongoing apps, you wouldn’t resend a whole library each time; cost and efficiency still matter.
• Many are excited about combining RAG + giant context: RAG to pick relevant chunks, big window to hold broad, multi‑source evidence for more reliable synthesis.

Privacy, AR & surveillance

• Long context plus AR sparks visions of “name tags over everyone” and personal-CRM overlays:

  • Seen as very useful for people who struggle with names/faces, and analogous to political “Farley files” and CRM systems.
  • Others call it a privacy nightmare: comprehensive behavioral memory erodes the social benefit of forgetting and invites stalking, blackmail, and honeypots.
  • Some think social norms would eventually adapt; others fear people will become ultra‑guarded or politician‑like in public.

• Separate thread imagines intelligence agencies (e.g., NSA) using long‑context LLMs to query massive surveillance archives:

  • Some claim NSA storage is more limited than people assume relative to big tech clouds.
  • Others cite Utah data center and large cloud contracts, doubting those reassurances and expecting extensive data plus AI analysis.

Energy, climate & resource concerns

• Several worry about data‑center power use and AI’s contribution to climate change, especially if much usage is trivial (e.g., cat pictures, low‑value chat).
• Others note compute is currently a small slice of global emissions compared to construction or travel, and big AI firms are relatively aggressive on renewables.
• There is skepticism about “carbon neutral” claims and carbon offsets; some insist we still need to minimize emissions rather than rely on accounting tricks.
• Agreement that AI demand is trending up and we may need better ways to assess whether its net benefits justify the resource cost.

Product quality, degradation & Google vs OpenAI

• Some fear public deployment will degrade Gemini’s quality (like perceived ChatGPT “nerfs”), and urge Google to prioritize consistency even if that means higher prices or limited access.
• Others question whether “degradation” is real or mostly perception plus occasional bad outputs.
• A few testers report Gemini Ultra is underwhelming for coding; others say Gemini 1.5 Pro feels fast and capable but is currently slow and lightly loaded.
• A minority dismiss Gemini outright because of Google’s perceived decline in search quality or ideological tilt.

Societal and labor impacts

• Mixed feelings about AI’s rapid progress:
  • Some are excited about productivity boosts (coding, law, summarization).
  • Others are anxious that elites will capture most benefits and that lower‑income people may be further marginalized, even to the point of “soft elimination” via economic pressure.
• A counterview expects broad price reductions and possible policy responses (e.g., stimulus, maybe UBI), though outcomes are acknowledged as uncertain.
• Analogy debates: will AI be like drum machines (augmenting, not replacing humans) or like cars vs carriages (wiping out old roles)? Many expect more complex, AI‑assisted work rather than outright disappearance of all jobs.