Phind-405B and faster, high quality AI answers for everyone

Original Article ↗ Hacker News Discussion ↗

Phind’s release of its Phind-405B model and revamped “Instant” search is prompting comparisons with tools like ChatGPT, Perplexity, Kagi, and Brave’s AI search, with many users praising Phind’s technical focus, coding help, and inline citations but criticizing UI glitches, occasional hallucinations, and a lack of API access. Some see it as a strong coding and research assistant—especially via its VS Code integration—while others have churned to competitors over reliability issues, subscription pricing, region blocks, or trust concerns. A recurring theme is the trade-off between speed, accuracy, and verifiability in AI-powered search, and how much users can safely rely on these systems without constant manual checking.

Usage patterns and strengths

  • Many use Phind as an AI-enhanced technical search engine, especially for programming, APIs, debugging, and infrastructure “how do I do X?” tasks.
  • Several report it as a strong productivity booster, getting them from near-zero knowledge (e.g., AWS VPC/NAT/Fargate) to working solutions quickly.
  • Common workflows: replacing Google + Stack Overflow; summarizing articles via URL; code optimization and debugging; niche language questions; learning new tech concepts.
  • Users like the presence of linked sources when they appear, treating Phind as “search + oracle” rather than pure chat.

Comparisons with competitors

  • Compared with ChatGPT, some prefer Phind for citations and technical focus, and as a fallback when ChatGPT has access/captcha issues.
  • Others prefer Kagi Assistant, Brave Search, Bing + GPT‑4o, Perplexity, or Claude for equal or better answers, broader features, or fewer UI issues.
  • Several note Phind-70B and now 405B can be competitive with Claude/GPT‑4 on some coding tasks, while GPT‑4 remains best for certain formatting tasks.

Hallucinations, accuracy, and verification

  • Multiple reports of confident but wrong answers: nonexistent language features, incorrect C++/Laravel examples, misdescribed hardware, and factual questions without valid references.
  • Users appreciate when Phind later admits a reference error or, in newer runs, detects nonsensical queries and corrects itself.
  • Some say “Always search” sometimes fails to trigger; others see answers improving when rerun.
  • General consensus: model is powerful but must be treated skeptically; follow‑up questions and checking sources remain essential.

Speed vs. quality

  • Thread discusses latency as a key barrier for AI search versus classic search.
  • Some argue that while token-by-token generation is slower, total “time to understanding” can be faster than traditional search, provided answers are accurate.

Product experience and UI

  • Positive: VS Code extension, “artifacts”-style features in development, improved search pollution, and better answer organization promised.
  • Negative: buggy web UI (scroll jumps, input obscured on mobile), occasional inference outages, region blocking (e.g., Malaysia), and some users being IP‑blocked.

Pricing, access, and “for everyone” claim

  • New Phind‑405B is only for paid Pro users; “for everyone” is interpreted by some as misleading marketing.
  • Phind Instant remains free; some want at least a small free quota for 405B to trial it.
  • Pricing criticized for having only a $20/month tier; some want cheaper, low‑usage plans.

API, ecosystem, and openness

  • Many request an API and OpenRouter‑style access so they can integrate Phind into their own tools and compare it on public leaderboards.
  • Company indicates API is lower priority than the main product but is now under consideration.
  • Some want weights released (especially Instant/70B), with debate over whether Llama’s license requires that; unclear from thread.
  • Concerns raised about opaque data handling and trustworthiness; one user says attempts to clarify for corporate use went unanswered.

Model behavior and philosophy

  • Long subthread critiques LLM “apologies” and anthropomorphic phrasing as misleading, since models lack real understanding, memory of wrongdoing, or capacity for genuine care.
  • Others stress that hallucinations and lack of “I don’t know” are structural to current LLMs; research on source‑aware training and better reasoning is referenced as a path forward.
  • Some propose using LLMs mainly to generate good search keywords and filter human-written sources, rather than as direct answer generators.