Gemini 3 Excels in Visual Creation but Lacks Agentic Reliability - Episode Hero Image

Gemini 3 Excels in Visual Creation but Lacks Agentic Reliability

This Day in AI Podcast · · Listen to Original Episode →
Original Title:

TL;DR

  • Gemini 3 Pro's significant improvement in visual generation and text rendering capabilities, particularly with Nano Banana Pro, enables the creation of complex infographics and marketing materials, potentially disrupting design software like Canva by offering AI-first, prompt-driven content generation.
  • Gemini 3 Pro's advanced image manipulation, including precise element editing and 4K upscaling, raises concerns about the trustworthiness of visual evidence, as it becomes increasingly difficult to distinguish AI-generated alterations from authentic images.
  • Gemini 3 Pro's "path obsession" and occasional context drift hinder its effectiveness in complex, multi-step tasks, requiring users to frequently reset or pivot, which diminishes its utility as a general-purpose agent compared to more grounded models.
  • Grok 4.1's exceptionally low pricing and robust tool-calling capabilities, combined with its strong citation referencing, position it as a highly cost-effective option for research and agentic tasks, despite potential trust issues and limited adoption.
  • The rapid advancement of multimodal AI models like Gemini 3 Pro and Nano Banana Pro suggests a future where specialized software may be supplanted by a universal AI assistant capable of generating custom UIs and executing complex creative tasks via simple prompts.
  • OpenAI's GPT-5.1 Pro, while demonstrating strong reasoning and instruction-following for complex problems, suffers from slow response times and a less integrated interface, making Gemini 3 Pro a more practical choice for daily creative and design-oriented workflows.
  • The increasing sophistication of AI image generation, particularly Gemini 3 Pro's ability to maintain character consistency across multiple inputs, challenges the notion of visual authenticity and necessitates a shift towards AI direction and verification skills for users.

Deep Dive

Google's Gemini 3 represents a significant, albeit uneven, leap in AI capabilities, excelling in creative coding and image generation while showing persistent weaknesses in core agentic functions like tool calling and instruction adherence. This release, alongside advancements in image generation with Nano Banana Pro, signals a strategic push towards multimodal applications, particularly in design and content creation, potentially disrupting established software markets. However, despite its impressive benchmarks and creative output, Gemini 3's struggles with instruction following and context drift suggest that while it offers a more refined user experience for specific creative tasks, it has not yet achieved the robust, general-purpose agentic performance required for complex, multi-step workflows, leaving room for competitors like Claude Haiku and Grok 4.1 to maintain an edge in more analytical and tool-centric applications.

Gemini 3's primary strength lies in its enhanced multimodal capabilities, especially in code generation and image creation, as demonstrated by its ability to produce complex 3D games and highly detailed infographics. This advancement is partly attributed to Google's integration of technologies like Winsurf and the development of the AnyGravity platform, which seem to have fine-tuned Gemini 3 for visual and a coding-centric output. The "Nano Banana Pro" image model, in particular, exhibits unprecedented text legibility and character consistency, enabling sophisticated use cases such as generating entire ad campaigns or highly detailed, contextually linked images. The implications are profound: businesses can leverage these tools for rapid content creation and marketing, potentially diminishing the need for specialized design software like Canva, which may see its market share eroded as AI-generated visuals become more accessible and sophisticated. This shift implies a future where user interaction with software becomes more conversational and directive, with AI adapting interfaces to specific tasks rather than users navigating predefined software structures.

However, Gemini 3's utility is tempered by significant second-order implications stemming from its persistent flaws, particularly in agentic behavior and tool use. Despite improvements, its tendency towards "path obsession" and context drift, where the model becomes fixated on a specific solution or deviates from instructions, necessitates constant human intervention. This makes it less reliable for complex, multi-stage tasks compared to models like Claude Haiku, which demonstrates superior groundedness and tool-calling capabilities. The "fatal patricia" persona shift, while an example of emergent creativity, highlights a potential lack of control and predictability, raising concerns about trustworthiness in critical applications. Furthermore, the pricing structure for Gemini 3, while not explicitly detailed, is implied to be higher, positioning it as a premium model. This, coupled with its limitations in agentic tasks, suggests that while Gemini 3 is a powerful tool for creative output, its value proposition for task-oriented AI agents remains incomplete, requiring users to potentially integrate multiple models or workarounds to achieve comprehensive workflow automation. Competitors like Grok 4.1, with its exceptionally low cost and strong tool-calling performance, and Claude Haiku, with its reliable and grounded agentic capabilities, present compelling alternatives for users prioritizing task completion and cost-effectiveness over cutting-edge creative generation.

Action Items

  • Audit Gemini 3 Pro's instruction following for path obsession by testing 5 complex, multi-step prompts and documenting instances of repeated solutions.
  • Evaluate Grok 4.1's tool-calling capabilities by designing 10 multi-tool tasks and comparing its success rate and efficiency against Gemini 3 Pro.
  • Implement a system to track Gemini 3 Pro's creative writing output quality by comparing 3-5 generated pieces against previous versions or other models.
  • Measure the impact of Gemini 3 Pro's visual generation on ad creation by generating 20 ad variants for a hypothetical product and assessing their quality and adherence to style prompts.
  • Analyze Nano Banana Pro's image manipulation accuracy by attempting 5 complex edits on original photos and documenting fidelity loss or unintended alterations.

Related Episodes

Key Quotes

"so chris this week it is finally here we have gemini 3 we also this morning got nano banana 2 that's why i'm wearing my yellow shirt and that's why we're recording two hours later than we planned because we've just been making images all morning yeah it is we've sunk a lot of time into it and probably a lot of money too we also got from xai grok 4 5 with a shocking 2 million context we'll talk about that a little bit later it's actually grok 4 1 but some people are saying it should be called 4 5 oh 4 5 did i say 4 5 i meant 4 1 we got gpt 5 1 codex max i'm not even kidding that's like a real model name and we got gpt 5 1 pro as well so we'll talk about those i don't think anyone really cares about it now everyone's really into the nano banana and the gemini 3 so we're going to talk about it we've also got a pretty good disc track and some other songs that we've created with gemini 3"

The speaker announces the release of several new AI models, including Gemini 3 and Nano Banana 2, highlighting the significant time and resources invested in exploring their capabilities. This quote sets the stage for a discussion of cutting-edge AI advancements and their immediate impact on image generation and model performance.

"i feel like this is at a minimum restored it to the level it was at before it's hard to say for me if it's better or not just yet but it does seem pretty good it's definitely faster which is a kind of nice benefit of it and i've also done some other testing i've actually done a bit of ai betting with it just to see how it performs and i've got some interesting results from that later so yeah it's interesting you say that because i think the benchmarks it's by far on the benchmarks the best model i think it lost to claude's sonnet 4 5 on one of the coding benchmarks but outside of that it's truly frontier on every account and you know so you look at those benchmarks and you think wow like it's you know it's really blasted ahead but then my actual experience using it and i suspect it's because similar to you i used gemini 2 5 pro a lot and my gut instinct feels like the reactions coming from a lot of people are because they probably never gave gemini 2 5 pro a shot"

The speaker offers an initial impression of Gemini 3, suggesting it has at least returned to the performance level of previous versions and is noticeably faster. They note that while benchmarks indicate it's a top-tier model, their personal experience suggests that some of the excitement might stem from users not having fully utilized Gemini 2.5 Pro previously.

"i feel like this is at a minimum restored it to the level it was at before it's hard to say for me if it's better or not just yet but it does seem pretty good it's definitely faster which is a kind of nice benefit of it and i've also done some other testing i've actually done a bit of ai betting with it just to see how it performs and i've got some interesting results from that later so yeah it's interesting you say that because i think the benchmarks it's by far on the benchmarks the best model i think it lost to claude's sonnet 4 5 on one of the coding benchmarks but outside of that it's truly frontier on every account and you know so you look at those benchmarks and you think wow like it's you know it's really blasted ahead but then my actual experience using it and i suspect it's because similar to you i used gemini 2 5 pro a lot and my gut instinct feels like the reactions coming from a lot of people are because they probably never gave gemini 2 5 pro a shot"

The speaker offers an initial impression of Gemini 3, suggesting it has at least returned to the performance level of previous versions and is noticeably faster. They note that while benchmarks indicate it's a top-tier model, their personal experience suggests that some of the excitement might stem from users not having fully utilized Gemini 2.5 Pro previously.

"i feel like this is at a minimum restored it to the level it was at before it's hard to say for me if it's better or not just yet but it does seem pretty good it's definitely faster which is a kind of nice benefit of it and i've also done some other testing i've actually done a bit of ai betting with it just to see how it performs and i've got some interesting results from that later so yeah it's interesting you say that because i think the benchmarks it's by far on the benchmarks the best model i think it lost to claude's sonnet 4 5 on one of the coding benchmarks but outside of that it's truly frontier on every account and you know so you look at those benchmarks and you think wow like it's you know it's really blasted ahead but then my actual experience using it and i suspect it's because similar to you i used gemini 2 5 pro a lot and my gut instinct feels like the reactions coming from a lot of people are because they probably never gave gemini 2 5 pro a shot"

The speaker offers an initial impression of Gemini 3, suggesting it has at least returned to the performance level of previous versions and is noticeably faster. They note that while benchmarks indicate it's a top-tier model, their personal experience suggests that some of the excitement might stem from users not having fully utilized Gemini 2.5 Pro previously.

"i feel like this is at a minimum restored it to the level it was at before it's hard to say for me if it's better or not just yet but it does seem pretty good it's definitely faster which is a kind of nice benefit of it and i've also done some other testing i've actually done a bit of ai betting with it just to see how it performs and i've got some interesting results from that later so yeah it's interesting you say that because i think the benchmarks it's by far on the benchmarks the best model i think it lost to claude's sonnet 4 5 on one of the coding benchmarks but outside of that it's truly frontier on every account and you know so you look at those benchmarks and you think wow like it's you know it's really blasted ahead but then my actual experience using it and i suspect it's because similar to you i used gemini 2 5 pro a lot and my gut instinct feels like the reactions coming from a lot of people are because they probably never gave gemini 2 5 pro a shot"

The speaker offers an initial impression of Gemini 3, suggesting it has at least returned to the performance level of previous versions

Resources

External Resources

Books

Videos & Documentaries

Research & Studies

Tools & Software

  • Suno - Used for creating and producing songs.

Articles & Papers

People

  • Sam Altman - Mentioned in relation to OpenAI and his company's strategy.
  • Elon Musk - Mentioned in relation to XAI, Grok, Tesla, and his public persona.
  • Greg Brockman - Mentioned in relation to a sad song he wrote.
  • Matt Schumer - Mentioned for his blog post and review of GPT-5.1 Pro.

Organizations & Institutions

  • OpenAI - Mentioned as a competitor in the AI model space.
  • XAI - Mentioned as the developer of Grok models.
  • Google - Mentioned as the developer of Gemini models and Nano Banana Pro.
  • Tesla - Mentioned in relation to Elon Musk's portfolio and potential AI applications.
  • Anthropic - Mentioned as a competitor in the AI model space, specifically regarding Claude models.
  • Nvidia - Mentioned in relation to its earnings and financial data.
  • Canva - Mentioned as a design tool potentially impacted by new AI image generation capabilities.
  • Microsoft - Mentioned in relation to its AI strategy and products.
  • Apple - Mentioned in relation to AI model negotiations.
  • CNBC - Mentioned as a potential outlet for news regarding AI model capabilities.
  • CNN - Mentioned as a potential outlet for news regarding AI model capabilities.
  • Motley Fool - Mentioned as a source for stock investment research.
  • Zacks - Mentioned as a source for stock investment research.
  • US News - Mentioned as a source for late 2025 projections and estimates.

Courses & Educational Resources

Websites & Online Resources

  • simtheory.ai - Website associated with the podcast hosts and their product.
  • X (formerly Twitter) - Mentioned as a platform for discussions and sharing information about AI models.

Podcasts & Audio

  • This Day in AI Podcast - The podcast where the discussion is taking place.

Other Resources

  • Gemini 3 Pro - An AI model discussed for its capabilities in text generation, coding, and image creation.
  • Nano Banana Pro - An AI image generation model discussed for its advanced capabilities in creating legible text, infographics, and detailed character consistency.
  • Grok 4.1 - An AI model from XAI discussed for its large context window, tool calling, and citation capabilities.
  • GPT-5.1 Codex Max - An AI model mentioned as an improved version of previous GPT-5-Codex models.
  • GPT-5.1 Pro - An AI model discussed for its reasoning capabilities and instruction following, though noted for its slowness.
  • Claude Sonnet 4.5 - An AI model mentioned in comparison to Gemini 3 Pro and Grok 4.1.
  • Claude Opus - An AI model mentioned in comparison to Gemini 3 Pro and Grok 4.1.
  • Claude Haiku - An AI model mentioned for its tool calling capabilities.
  • Gemini 2.5 Pro - A previous version of the Gemini model discussed in comparison to Gemini 3 Pro.
  • Gemini 2.5 Flash - An earlier image model from Google, previously known as Nano Banana.
  • Constitutional AI - A safety framework mentioned in relation to Anthropic's models.
  • RAG (Retrieval-Augmented Generation) - A technique mentioned in the context of AI models finding context across files.
  • Synth ID watermark - A watermark technology for AI-generated images.
  • V8 engine visualization - An example of a complex visualization created with AI.
  • Lunar lander game - An example of a game created with AI.
  • 3D village game with Santa Claus - An example of a game created with AI.
  • Plant cell diagram - An example of an educational diagram created with AI.
  • System diagram for security - An example of a technical diagram created with AI.
  • AI betting experiment - An experiment using AI models for sports betting analysis.
  • Fatal Patricia - A persona adopted by the AI model Patricia.
  • Jewels - An AI agentic coding tool from Google.
  • Antigravity - A product from Google that includes an IDE and VS Code fork.
  • AI Studio - A Google product for AI development.
  • NotebookLM - A Google product for AI-powered note-taking and research.
  • Vertex AI - A Google Cloud AI platform.
  • VS Code - A code editor mentioned in relation to Antigravity.
  • Photoshop - A photo editing software mentioned for comparison with AI image manipulation.
  • Commodore 64 - An older computer mentioned for nostalgic reasons.
  • Calculator - A tool mentioned in comparison to AI's ability to perform calculations.
  • AI Search - AI integration within search engines.
  • Agentic loop - The process of AI agents performing tasks and iterating.
  • Tool calling - The ability of AI models to use external tools.
  • Context drift - The tendency of AI models to lose track of the original context.
  • Path obsession - The tendency of AI models to get stuck on a particular solution path.
  • Instruction following - The ability of AI models to adhere to user instructions.
  • Multimodal world - A world where AI can process and generate various types of data (text, images, audio, video).
  • UI (User Interface) - The visual elements of a digital product.
  • SAS (Software as a Service) - A software distribution model.
  • B2B SAS - Business-to-business software as a service.
  • Simlink - A future product or feature related to agentic tasks.
  • Patreon - A platform for creators to receive financial support from fans.
  • Spotify - A music streaming service.
  • Discord - A communication platform.
  • LinkedIn - A professional networking platform.
  • Google Photos - A photo storage and management service.
  • Wedding invite maker pro - A hypothetical specialized tool.
  • Slide presentation pro - A hypothetical specialized tool.
  • CSI Miami interface - A visual interface style.
  • Meta's Segment Anything Model - An AI model for image segmentation.
  • Open-source models - AI models with publicly available code.
  • Open-weight models - AI models with publicly available weights.
  • Grok paper - The documentation or research paper for the Grok model.
  • Chemical weapons - Mentioned as an example of content that might be restricted by AI safety filters.
  • Weird sex stuff - Mentioned as an example of content that might be restricted by AI safety filters.
  • Australian laws - Mentioned in relation to potential legal ramifications of AI image manipulation.
  • AI image verification - Technology to check the authenticity of AI-generated images.
  • Google Adwords - Google's advertising platform.
  • Tiktok style frame - A visual style for content.
  • California style - A visual style.
  • Human eggs billboard - An example of surrealist AI-generated content.
  • Horse eggs - An example of surrealist AI-generated content.
  • Huntsman spider - A type of spider mentioned in an image manipulation example.
  • Kangaroo - An animal mentioned in an image manipulation example.
  • Passport image - Mentioned in the context of potential photo forgery.
  • Insurance claim on accident photo - Mentioned in the context of potential photo forgery.
  • Slandering someone in the newspaper - Mentioned in the context of potential photo forgery.
  • AI studio - A platform for creating AI applications.
  • AI model editing - The process of modifying AI models.
  • AI model tuning - The process of adjusting AI models for better performance.
  • AI model fine-tuning - The process of adapting pre-trained AI models to specific tasks.
  • AI model training - The process of developing AI models.
  • AI model release - The launch of new AI models.
  • AI model benchmarks - Standardized tests to evaluate AI model performance.
  • AI model pricing - The cost associated with using AI models.
  • AI model capabilities - The range of tasks an AI model can perform.
  • AI model limitations - The constraints of an AI model's performance.
  • AI model hallucinations - Instances where AI models generate incorrect or fabricated information.
  • AI model creativity - The ability of AI models to generate novel and imaginative content.
  • AI model reasoning - The ability of AI models to process information and draw conclusions.
  • AI model safety - Measures to ensure AI models operate ethically and responsibly.
  • AI model censorship - Restrictions placed on the content AI models can generate.
  • AI model performance - How well an AI model executes its tasks.
  • AI model development - The process of creating and improving AI models.
  • AI model integration - The process of incorporating AI models into existing systems.
  • AI model deployment - The process of making AI models available for use.
  • AI model research - The study and exploration of AI models.
  • AI model applications - The practical uses of AI models.
  • AI model advancements - Progress and improvements in AI models.
  • AI model evolution - The ongoing development and changes in AI models.
  • AI model impact - The effects of AI models on society and industries.
  • AI model future - Predictions and expectations for the future of AI models.
  • AI model ecosystem - The network of AI models, tools, and platforms.
  • AI model landscape - The current state and trends in AI models.
  • AI model innovation - The introduction of new ideas and technologies in AI models.
  • AI model disruption - The impact of AI models on existing industries and business models.
  • AI model competition - The rivalry between different AI model developers.
  • AI model adoption - The rate at which AI models are being used.
  • AI model accessibility - The ease with which users can access and utilize AI models.
  • AI model user experience - The overall interaction users have with AI models.
  • AI model user interface - The visual elements through which users interact with AI models.
  • AI model user feedback - Input and opinions from users about AI models.
  • AI model user satisfaction - The level of contentment users have with AI models.
  • AI model user engagement - The extent to which users interact with AI models.
  • AI model user behavior - The patterns of how users interact with AI models.
  • AI model user needs - The requirements and desires of users for AI models.
  • AI model user preferences - The choices users make when interacting with AI models.
  • AI model user expectations - What users anticipate from AI models.

  • AI model user trust - The confidence users have in AI models.

---
Handpicked links, AI-assisted summaries. Human judgment, machine efficiency.
This content is a personally curated review and synopsis derived from the original podcast episode.