Bard and PaLM 2: Google’s Bet on Generative AI

javi
Pantalla con asistente de IA conversacional

Google launched Bard in February 2023 as a response to ChatGPT’s cultural impact, and in May presented PaLM 2 as the model powering it. After three months of intensive industry use, the differences between Bard+PaLM 2 and GPT-4 are clearer: it’s not a race for who generates better text but for who integrates generative AI best into the rest of their ecosystem.

PaLM 2 in Context

PaLM 2 is the second generation of Google’s Pathways Language Model. Published with a very condensed technical paper (no model size, no detailed training dataset), we know:

  • Ships in four sizes (Gecko, Otter, Bison, Unicorn). Bard uses Unicorn; developers access Bison via Vertex AI.
  • Trained with a higher share of multilingual, code, and scientific data than PaLM 1.
  • Competes with GPT-3.5/GPT-4 on standard benchmarks: MMLU, BIG-bench, Winograd. Results oscillate — it wins some, loses some.

In real applications, PaLM 2 in Bard is solid for:

  • Summarising and rewriting long text. The expanded context window (up to 38k tokens on Bison) makes work on long documents easier.
  • Search with reasoning. Bard has native access to Google Search — it can cite recent sources, something ChatGPT (without browsing) can’t.
  • Quality multilingual. Particularly in less common language pairs, PaLM 2 gives more consistent results than other large models.

Where GPT-4 Still Leads

Two areas where GPT-4 maintains a clear edge in July 2023:

  • Complex step-by-step reasoning. On competition math, symbolic logic, or detailed legal reasoning, GPT-4 shows higher reliability.
  • Code generation and refactoring. GPT-4’s integration with GitHub Copilot has matured faster than Google’s proposition — though Codey is starting to close the gap.

Developer ecosystem also weighs heavily: after six months, there’s an order of magnitude more libraries, tutorials, and products built on OpenAI’s API than on Vertex AI.

Google’s Strategic Advantage

Despite PaLM 2 not beating GPT-4 on average, Google holds a unique card: suite integration. In May Google announced “Duet AI” for Workspace — text generation in Gmail, auto-summaries in Meet, drafts in Docs, formulas in Sheets. For a company already on Workspace, this requires no integration: it appears where the team already works.

Microsoft does the same with Microsoft 365 Copilot on GPT-4. The real 2024 fight won’t be “which model is better” but “which productivity suite integrates it most usefully”.

Bard as a Product

Bard itself (chat.openai.com/chat vs bard.google.com) lives in the background. Google uses it as direct consumer touchpoint and as a channel to experiment with features:

  • Share conversations via public link.
  • Gmail/Docs integration to export generated content directly to an email or document.
  • Colab export for data and code.

Bard isn’t yet a direct competitor to ChatGPT Plus on price (free) or plugins (limited compared to OpenAI’s ~500), but each update closes some of the gap.

Gemini on the Horizon

Google’s next piece is Gemini, announced at I/O 2023 as the fusion of Google Brain and DeepMind teams to produce a native multimodal model. According to leaks and statements from Sundar Pichai and Demis Hassabis, Gemini will be trained on a much larger dataset and designed with tool-use (action, not just text) in mind.

Late-2023 bets: Gemini could be the moment Google perhaps surpasses GPT-4 on raw capability. If that holds, the “best model + best integration” architecture could be difficult to counter.

Practical Considerations

For teams evaluating which LLM to bet on:

  • Already on Workspace or GCP: Vertex AI with PaLM 2 has the lowest-friction natural integration.
  • Already on Azure or AWS: OpenAI via Azure OpenAI Service or Bedrock gives more model options (GPT-4, Claude, Llama).
  • New decision: test both with prompts from your real use case. Academic benchmarks rarely predict behaviour in your domain.

Also see ChatGPT 4 advances as practical-comparison reference.

Conclusion

Bard with PaLM 2 doesn’t beat ChatGPT with GPT-4 on raw capability, but Google is playing a different game: winning by integration within the Workspace ecosystem. For 2024, with Gemini available, the race could rebalance — but the real fight will be who offers the best end-to-end experience, not who has the biggest model.

Follow us on jacar.es for more on LLMs, enterprise AI, and platform strategy.

Entradas relacionadas