Hi everyone,
With the ongoing rollout of Gemini 3.5 Flash and the anticipated release of Gemini 3.5 Pro, I want to open a discussion regarding Google’s next-generation model strategy—specifically how future architectures will stack up against upcoming leaps from Anthropic and OpenAI when it comes to deep, deterministic reasoning.
I am currently developing large-scale Battery Energy Storage Systems (BESS). My workflow requires AI models to strictly reason based on physical laws, electrical engineering principles, and highly specific internal development standards.
When using current conversational models, I face two massive roadblocks that make me question if current LLM baselines are enough, or if we genuinely need to train/fine-tune specialized engineering models:
- Physical Blindspots: Standard conversational LLMs frequently hallucinate or misinterpret strict thermodynamic and electrical constraints.
- Context Amnesia: Because models “forget” everything after a session change, I am forced to bootstrap every single interaction by pasting massive amounts of boilerplate code, proprietary “internal laws,” and development frameworks just to get the model on the same page.
To cope with this, I’ve had to build a makeshift “work storage system” to force compliance, but it is highly inefficient for complex industrial engineering.
As we look toward the next generation of models that aim to compete with GPT-Next or Claude 4, I would love to hear from the community or Google Developer Advocates on the following:
- Advanced Reasoning vs. Hard Physics: Will the upcoming generation feature breakthroughs in grounding models to actual physical and mathematical laws, or will they remain purely language-probabilistic?
- State and Memory Management: Are there native architectural upgrades planned to solve “context amnesia” for developers who need persistent, long-term alignment with massive codebases and proprietary standards without burning through context windows on every reset?
- The Engineering Path: For complex fields like BESS, is Google’s roadmap steering toward highly capable general models, or should we actively pivot toward fine-tuning open-source models with dedicated engineering datasets?
Looking forward to any insights, official documentation, or discussion on how the next wave of Gemini will tackle these deep-tech challenges!