OpenAI Trapped in 2nd place - CourtneyR.dev's 2nd Brain

OpenAI: Trapped in 2nd place 🔗 **Source:** [OpenAI: Trapped in 2nd place](https://www.youtube.com/watch?v=_nxbZgZysT4) > [!note]+ Document Notes > tags: [ai-industry, openai, market-competition, chatgpt, ai-models] > author: Theo - t3․gg > source: YouTube > reference: https://www.youtube.com/watch?v=_nxbZgZysT4 > date created: 2025-01-25 13:30 > > # 🏷️ Title: OpenAI: Trapped in 2nd Place >  ## 🧰 Action Inbox - [H1] Evaluate OpenAI's strategy vs. competitors across programming use cases — 03:45 #programming - [H2] Monitor chat experience market share changes for competitive threats — 15:20 #strategy - [H3] Track document analysis capabilities gap between models — 25:10 #capabilities ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [OpenAI Models](https://platform.openai.com/docs/models) - [Anthropic Claude](https://www.anthropic.com/claude) ## 📝 Summary - OpenAI consistently ranks second across programming, image generation, document analysis, and tool calling despite releasing groundbreaking features first - Their primary competitive advantage remains the ChatGPT interface, which captures 70% of their revenue and maintains user retention - Google and Anthropic repeatedly surpass OpenAI's technical capabilities within months of major releases - OpenAI's strategy focuses on "good enough" performance across categories rather than excellence in any single domain - The company's survival depends on maintaining chat experience leadership rather than technical superiority ## 📖 Main Content OpenAI faces a recurring pattern where they introduce revolutionary AI capabilities only to be surpassed by competitors within months. Despite pioneering reasoning models with GPT-5, DeepSeek's R1 matched performance while offering open weights. Similarly, GPT Image dominated until Google's "Nano Banana" models exceeded capabilities at faster speeds. The analysis examines six key AI categories: programming, chat experience, image generation, document analysis, agent tool calling, and browser use. OpenAI ranks second or lower in five categories, with their only first-place position being chat experience through ChatGPT. This positioning reflects OpenAI's strategic focus on platform retention rather than technical dominance. With 70% of revenue from ChatGPT subscriptions, they prioritize maintaining user engagement over benchmark leadership. The approach succeeds because users prefer integrated "good enough" solutions over switching between specialized tools. However, this strategy carries risks. Google's distribution advantages and potential interface improvements could threaten OpenAI's chat dominance. The company's "Code Red" response to a 7% traffic decline demonstrates their sensitivity to user retention metrics. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/_nxbZgZysT4?t=0) — [H1] OpenAI's recurring competitive pattern - Key idea: OpenAI introduces paradigm-shifting features then gets surpassed by competitors - Evidence: "GPT-5 changed how models work entirely by introducing reasoning... DeepSeek drops R1 with similar numbers" #### ⏱️ [02:15](https://youtu.be/_nxbZgZysT4?t=135) — [H2] Market collapse from Chinese competition - Key idea: DeepSeek's open-weight model caused immediate market disruption - Evidence: "The whole market collapsed as a result" of free alternative with similar performance #### ⏱️ [03:45](https://youtu.be/_nxbZgZysT4?t=225) — [H1] Programming model rankings established - Key idea: Anthropic's Opus 4.5 currently leads programming tasks over GPT models - Evidence: "Opus is built different... it's my default now in cursor, shipping way more code" #### ⏱️ [06:30](https://youtu.be/_nxbZgZysT4?t=390) — [H2] CLI tool competition dynamics - Key idea: OpenAI's Codecs fell behind despite promising open-source start - Evidence: "Codecs has not meaningfully improved... Cloud Code has found its position pretty far ahead" #### ⏱️ [09:15](https://youtu.be/_nxbZgZysT4?t=555) — [H1] Chat experience as OpenAI's stronghold - Key idea: ChatGPT interface remains OpenAI's primary competitive advantage - Evidence: "70% of their revenue... when people refer to AI, they're often referring to ChatGPT" #### ⏱️ [12:20](https://youtu.be/_nxbZgZysT4?t=740) — [H2] Model conversation quality assessment - Key idea: GPT models produce unpleasant emoji-heavy responses compared to competitors - Evidence: "I am so tired of the emoji riddled slop that we get out of GPT" #### ⏱️ [15:45](https://youtu.be/_nxbZgZysT4?t=945) — [H1] Image generation leadership shift - Key idea: Google's Nano Banana models surpassed GPT Image in speed and capability - Evidence: "Nano Banana generates in seconds... no piss filter, roughly same accuracy" #### ⏱️ [19:30](https://youtu.be/_nxbZgZysT4?t=1170) — [H2] Nano Banana Pro's advanced capabilities - Key idea: Google's model generates complete slide decks with proper text integration - Evidence: "It can generate full slide decks... proper spatial awareness... almost scary" #### ⏱️ [22:10](https://youtu.be/_nxbZgZysT4?t=1330) — [H1] Document analysis dominance by Google - Key idea: Gemini models excel at large context document processing - Evidence: "Gemini is so far ahead... chews through PDFs, finds needles in haystacks really well" #### ⏱️ [25:15](https://youtu.be/_nxbZgZysT4?t=1515) — [H2] Tool calling reliability comparison - Key idea: Anthropic leads in reliable agent tool usage over OpenAI - Evidence: "Opus 4.5 closed whatever gaps... I don't get errors from Opus, occasionally get errors from GPT" #### ⏱️ [28:45](https://youtu.be/_nxbZgZysT4?t=1725) — [H1] Scoring system reveals OpenAI strategy - Key idea: Despite second-place rankings, OpenAI leads overall through consistent performance - Evidence: Point scoring shows OpenAI at 12 points vs competitors at 10-11 #### ⏱️ [31:20](https://youtu.be/_nxbZgZysT4?t=1880) — [H1] "Good enough" strategy explanation - Key idea: OpenAI prioritizes platform retention over technical excellence - Evidence: "They don't have to be as good as Opus... if they're 5% worse but you're already paying, who cares?" #### ⏱️ [34:15](https://youtu.be/_nxbZgZysT4?t=2055) — [H2] Revenue model dependency on chat experience - Key idea: ChatGPT subscriptions drive 70% of revenue, making chat experience critical - Evidence: "Code Red happened because they saw 7% dip in week-over-week traffic" #### ⏱️ [36:50](https://youtu.be/_nxbZgZysT4?t=2210) — [H1] Google's distribution advantage threat - Key idea: Google's app ecosystem could threaten OpenAI if properly integrated - Evidence: "Google has insane distribution... if they integrate AI properly, incentive to use ChatGPT goes down" #### ⏱️ [38:30](https://youtu.be/_nxbZgZysT4?t=2310) — [H1] OpenAI's winning position despite second place - Key idea: Market leadership in chat experience ensures success regardless of technical rankings - Evidence: "They already won first in the only category that matters, the app you open when you want to talk to an AI" ## 📚 References - [OpenAI Platform Documentation](https://platform.openai.com/docs) - [Anthropic Claude Models](https://www.anthropic.com/claude) - [Google Gemini](https://gemini.google.com/) ## 🗣️ Paraphrase - OpenAI consistently introduces groundbreaking AI capabilities but gets surpassed by competitors within months, despite maintaining strong market position through ChatGPT - The company's "good enough" strategy across multiple AI categories generates more revenue than technical excellence in any single domain - Google poses the primary threat due to distribution advantages, but poor software execution currently limits their competitive impact ## 📰 Newsletter The AI landscape reveals a fascinating paradox: OpenAI keeps losing technical races while winning the market war. Despite ranking second in programming, image generation, and document analysis, their ChatGPT platform captures 70% of revenue through superior user experience. This "good enough" strategy proves more valuable than benchmark leadership. Users prefer integrated solutions over switching between specialized tools, even when competitors offer superior capabilities. OpenAI's real competition isn't technical—it's platform retention. For developers and businesses, this highlights the importance of distribution and user experience over pure performance metrics. The lesson? Sometimes being universally good beats being specifically great. Resources: [OpenAI Platform](https://platform.openai.com/), [Anthropic Claude](https://www.anthropic.com/claude), [Google AI](https://ai.google/) ## ✅ Action Items #### DevRel - Developer advocacy: Compare model performance across programming tasks, create benchmark reports - Education: Workshop on choosing AI models for different use cases, platform integration training - Developer experience: Integrate multiple AI providers, reduce vendor lock-in concerns - Developer marketing: Position multi-model strategies, emphasize workflow over individual model performance - Community management: Survey developer tool preferences, collect integration feedback #### OSPO - Policy & Governance: AI model usage policies, vendor evaluation criteria - License Compliance: Review open-weight model licenses, commercial usage terms - Security Management: Multi-vendor AI security assessments, data handling policies - Software Inventory: Catalog AI model dependencies, track usage patterns - Developer Support: Training on model selection, cost optimization strategies - Community Engagement: Fund open-source AI tooling, model evaluation frameworks - Strategic Alignment: KPIs for AI adoption, ROI measurement across providers #### Community - Maintainers: AI integration roadmaps, model selection guidance for projects - Contributors: Training on AI-assisted development, model comparison workshops - Sponsors: Transparent AI tool budgets, performance metrics reporting - Ecosystem: Cross-platform AI integration standards, vendor neutrality principles #### WordPress - Core: AI integration strategy, model selection for different core features - Core Performance: AI-assisted optimization, model performance impact analysis - Core AI: Multi-model evaluation, fallback strategies, cost management - Core Program Team: AI vendor coordination, integration dependencies, retrospectives - Mobile: AI feature parity, model optimization for mobile constraints - Playground: Multi-model experimentation, comparative demos, educational content - Community: AI tool recommendations, integration workshops, vendor comparison guides - Training: Model selection curricula, cost-benefit analysis lessons, integration paths - Marketing: Multi-model campaign strategies, vendor-neutral messaging - Support: AI troubleshooting guides, model-specific help documentation - Design: AI-assisted design workflows, tool integration UX - Documentation: Model comparison guides, integration documentation, best practices - Meta: AI infrastructure strategy, vendor management, cost optimization - CLI: Multi-model command support, vendor-agnostic interfaces - Hosting: AI model hosting recommendations, integration guides - Plugins: AI-powered plugin development, model selection guidelines ## 🧪 Hypotheses - User retention in AI platforms correlates more strongly with interface quality than model performance benchmarks - Companies achieve higher revenue through "good enough" multi-capability platforms than specialized best-in-class tools - Distribution advantages (existing user base, app ecosystem) matter more than technical superiority in AI adoption - Market leaders can sustain competitive positions despite consistent second-place technical rankings ### 📊 Methodology - Comparative analysis of AI model performance across programming, chat, image generation, document analysis, and tool calling categories - Revenue correlation study between platform features and technical benchmarks - User retention analysis comparing integrated platforms versus specialized tools ### ❓ Questions - How long can OpenAI maintain chat experience leadership against Google's distribution advantages? - What performance gap threshold would trigger significant user migration from integrated platforms? - How do enterprise adoption patterns differ from consumer preferences in AI tool selection? ### 🔮 Future Research - Quantitative analysis of user switching behavior between AI platforms - Enterprise decision-making factors in AI vendor selection - Distribution platform integration impact on AI adoption rates ### 💡 Implications - Practice: Prioritize platform integration and user experience over pure performance optimization - Policy: Develop vendor-neutral AI policies to avoid lock-in with technically superior but potentially unstable providers - Security: Implement multi-vendor strategies to reduce dependency on single AI providers - Education: Train teams on platform thinking rather than model-specific optimization - Funding: Allocate budgets for platform development over individual model licensing ### 🧠 Learning Styles Assessment Reading level estimate: Flesch-Kincaid Grade Level 8-9 #### 👀 Visual learners - Present: Performance comparison charts, model ranking tables - Improve: Interactive benchmark dashboards, visual model capability maps, competitive landscape diagrams #### 🔊 Aural learners - Present: Video format with clear narration, conversational explanations - Improve: Audio summary of key competitive insights, podcast-style deep dives #### 🗣️ Verbal learners - Present: Clear definitions of AI categories, analogies to traditional software markets - Improve: Glossary of AI model terminology, competitive analysis frameworks #### 🧩 Kinesthetic learners - Present: Model testing examples, hands-on performance comparisons - Improve: Interactive model comparison tools, step-by-step evaluation guides #### 🧠 Logical learners - Present: Point-scoring system, systematic category analysis - Improve: Decision trees for model selection, quantitative performance metrics #### 🧑‍🤝‍🧑 Social learners - Present: Community feedback integration, user experience sharing - Improve: Discussion forums for model comparisons, peer review platforms #### 🧘‍♀️ Solitary learners - Present: Individual analysis framework, self-directed exploration - Improve: Personal model evaluation checklists, reflection prompts on tool selection #### 🌿 Naturalistic learners - Present: Market ecosystem thinking, competitive landscape evolution - Improve: AI market lifecycle analogies, ecosystem impact analysis ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive section labels - Improve: Alt text for competitive comparison charts, table headers for model rankings #### Keyboard navigation - Present: Standard video controls, sequential content flow - Improve: Keyboard shortcuts for section navigation, accessible chart interactions #### Contrast and readability - Present: High contrast text, clear section separation - Improve: Color-blind friendly chart colors, adjustable contrast settings #### Text resizing and reflow - Present: Responsive text layout, scalable content - Improve: Text size controls, reflowable comparison tables #### Media and captions - Present: Video with spoken content - Improve: Comprehensive captions, downloadable transcript, audio descriptions #### Forms and interactions - Present: Basic video interaction controls - Improve: Accessible model comparison tools, form-based evaluation interfaces #### Standards reference - Map video content accessibility to WCAG 2.2 AA guidelines for media ### 🌍 DEIB Assessment #### Cultural breadth - Present: Focus on US-based AI companies, English-language platforms - Improve: Global AI market analysis, international model providers, localization considerations #### Inclusive language - Present: Technical jargon with explanations, industry-standard terminology - Improve: Plain --- ## Full Transcript [00:00] Open AI is a fascinating company. One [00:01] day they'll be introducing some crazy [00:03] paradigm that changes what we use AI for [00:05] and then the next day some small company [00:07] is trouncing them with the same [00:08] technology. I fondly think back to the [00:11] end of last year when 01 changed how [00:13] models work entirely by introducing [00:15] reasoning to the masses. This idea that [00:17] a model can generate tokens to steer [00:19] itself in the right direction and lead [00:21] itself to a more likely correct answer. [00:23] Just a few months later, not even like a [00:26] month and a half later, Deepseek drops [00:28] R1 with the same technique. But unlike [00:30] OpenAI, they actually share those [00:31] reasoning tokens because they shared an [00:33] openw weight model. You can use it for [00:35] whatever. They got similar numbers to [00:37] what OpenAI was getting and this thing [00:38] you could use for free. And immediately [00:40] like the whole market collapsed as a [00:42] result. This seemed like a one-off [00:44] thing, this rare moment where a small [00:46] company in China was able to copy, [00:48] reproduce, and exceed the research [00:50] OpenAI had done. But that continues to [00:53] happen and here we are today right after [00:55] GPT 5.2 dropped and Gemini 3 Pro is [00:58] already ahead and now Gemini 3 Flash is [01:01] right behind. I was so blown away with [01:03] the launch of GPT5 that I was using it [01:05] for as much as I could, which was [01:06] admittedly not too much cuz it was so [01:08] slow, but I still liked it a lot for [01:10] code and UI stuff. And just a few months [01:12] later, other models got better at UI and [01:14] are still faster and more reliable. So I [01:16] found myself using it less. And then [01:18] there's everything from codecs to their [01:20] web app to their image models and so [01:22] much more. It seems like OpenAI is [01:24] caught in this perpetual cycle where [01:26] they do something incredible, get [01:28] trumped, and now they're just second [01:30] place for a long time. I want to talk [01:32] more about this because I think it [01:34] showcases how OpenAI is different from [01:36] other companies, but also how tough it [01:38] is to compete in this market as a whole. [01:40] And alongside that, you might learn a [01:41] bit about how I think about these models [01:43] and labs and how I make decisions about [01:44] what I am using. As always, none of the [01:46] companies we're talking about now are [01:48] paying me, and I am hopeful, fingers [01:50] crossed, this isn't going to get me cut [01:51] off of the nice list with OpenAI, where [01:53] I get early access to things. Either [01:55] way, someone's got to pay us. So, we'll [01:56] do a quick break for today's sponsor. [01:58] These guys are all about saving time. [01:59] So, if you're not hiring anybody, you [02:01] can skip this ad. But if you are, you [02:03] really should check out today's sponsor [02:04] because G2I will make hiring so much [02:07] easier, so much faster, and probably [02:09] bring you better engineers than you [02:10] would otherwise get yourself. Their [02:12] network of 8,000 engineers is [02:13] unbelievable. They're not just a bunch [02:15] of fresh grads trying to get their first [02:16] gig. It's people who have had years upon [02:19] years of experience doing real projects [02:21] at real companies and their knowledge is [02:23] up to the vast majority of them are [02:25] already on board on the best AI tools. [02:27] So you're not going to have to teach [02:27] them how to use cursor. You can just [02:29] throw them at your codebase and get them [02:30] going. What's even crazier is how fast [02:32] they'll start filing those PRs. It's [02:34] G2's goal to get you from signed up to [02:36] PR filed within 7 days. Yes, really. You [02:40] sign up day one, you meet with their [02:41] team day two, they set up the Slack day [02:43] three, and they have candidates ready to [02:45] start on a trial by the end of the week, [02:47] and they will have PRs filed by the end [02:49] of that day. It's kind of crazy, but [02:50] I've seen it happen enough times, I do [02:52] actually believe it. If you're running a [02:54] team, you have way more important stuff [02:55] to be doing than chasing down [02:57] recruiters. Use the ones who get it at [02:59] soy. [03:01] OpenAI can't get out of second place. [03:03] The nose is a bold statement, but I want [03:05] to try and break down the different [03:07] categories of things we do with AI so I [03:09] can more meaningfully establish this as [03:12] obvious. There are things like [03:13] programming and I could break this down [03:15] to front end versus backend and other [03:17] things, but I'll keep it generic and [03:18] just put programming as a category. I'll [03:21] have like chat experience for people [03:23] using these models to like get homework [03:25] help, talk about their feelings, all the [03:27] things that your average normie would do [03:29] in chat GPT. I'll throw in some other [03:32] fun things like image generation. I [03:34] don't know what else do we got. Document [03:37] analysis. [03:39] Sure. Agent tool calling. [03:43] Browser use [03:45] can put in there. I'm not trying to say [03:47] any of these categories are equal or [03:48] that like programming is as big of a [03:50] deal as browser use and document [03:51] analysis. I'm just trying to break down [03:53] categories where having the best option [03:56] means something. Being 1% better at [03:58] translations for 5% more money doesn't [04:01] mean anything. Like everything's good [04:02] enough at translating now. But yeah, I [04:05] want to focus on a set of categories. [04:08] Actually, deep research is an [04:10] interesting one that I definitely will [04:12] get to later. But for now, let's focus [04:15] on these guys. We'll start with [04:17] programming. I don't think this is too [04:19] hot of a take at the moment, but in my [04:21] opinion, the best programming model by [04:23] far right now is Opus 4.5. I love [04:28] [ __ ] on anthropic more than almost [04:30] anyone does, but Opus is built [04:33] different. I will say that for some [04:36] really hard tasks that require a lot of [04:38] thinking and understanding of your [04:39] codebase, GBT 5.2x high can do slightly [04:43] better, but ends up doing way more [04:46] reasoning, way more tokens, taking [04:47] forever because it's so slow, and still [04:50] having the quirks of a GPT model versus [04:52] an Opus one where it just doesn't use [04:54] the harness quite as well. loves going [04:56] rogue and doing strange [ __ ] which is [04:58] weird because GBT 5 was so good at [04:59] destruction following before. I could [05:02] put multiple other anthropic and GPT [05:04] models in here, but I'm really trying to [05:06] think of this as the providers positions [05:07] in the list. And honestly, Gemini 3 Pro, [05:12] this does not properly represent the gap [05:14] I perceive between these things. I don't [05:16] like coding with Gemini. There's a lot [05:18] of reasons for it. It's a combination of [05:19] the model just going rogue and doing its [05:21] own [ __ ] not handling tool calls very [05:23] well. It's slightly better at UI than [05:25] the other two hypothetically. Sometimes [05:28] it's good at spatial reasoning. So if [05:29] you're making a 3D game and want the [05:31] model to place things for you, Gemini 3 [05:33] Pro is solid. But actually using it for [05:36] code on a day-to-day basis, not great. [05:38] 5.2 I don't like using for traditional [05:42] coding tasks, but if I'm planning a big [05:44] overhaul or like want to make sure this [05:46] bug is solved properly, I found 5.2 to [05:49] be really nice for debugging type tasks. [05:52] And then Opus as the general end all [05:54] beall just using it for everything all [05:55] of the time. It's my default now. I use [05:57] it all the time in cursor. I've been [05:59] shipping way more code as a result of [06:00] how much fun I'm having with Opus. It's [06:03] crazy how quickly this gap collapsed. [06:06] GPT5 was far ahead of everything else. [06:08] So much so that we dealt with its [06:09] quirks. We dealt with its slow speeds. [06:11] We dealt with its weird tool call [06:13] formatting. And now Opus is ahead again. [06:15] It barely matters. They keep fighting to [06:17] push it back. Like the 5.1 and 5.2 two [06:19] drops both felt like attempts to swing [06:21] back and take some of the win from [06:23] Anthropic and Google. They succeeded in [06:25] the case of Google in my opinion. I also [06:26] don't think that Gemini 3 it is very [06:29] smart the model knows a lot but I don't [06:31] think it's a great experience for [06:32] actually using. So very quickly GPT went [06:36] from the best option with five which is [06:38] still a crazy drop and it collapsed [06:40] quickly after. There's one other [06:42] programming category I want to put down [06:44] here just to continue leaning into this. [06:47] I'll call this CLI tool. It was really [06:51] cool to see OpenAI open sourcing codecs. [06:54] I was genuinely hyped that a big lab [06:57] took a thing as important as a CLI tool [06:59] for doing agent coding and rather than [07:01] hiding it behind closed doors and [07:03] sharing a binary that they accidentally [07:05] attach source maps to and then DMCA you [07:07] for sharing. Sorry, Enthropic, you're [07:09] you're winning me over on the models. I [07:10] do not like how you run cloud code [07:12] still. Codeex coming out as true open- [07:15] source was huge and I was genuinely [07:18] really excited to see it but they also [07:20] rewrote it in Rust which they did not [07:22] need to do. I would argue that when [07:24] Codeex dropped it was slightly ahead of [07:28] Cloud Code. It had roughly the same [07:29] functionality, way more extensibility. [07:32] The open- source part meant you could do [07:34] really cool [ __ ] with it. And at the [07:36] time, Cloud Code had somewhat stagnated [07:38] due to the leads for the project going [07:40] to cursor and then quickly boomeranging [07:42] back over to work at Claude Code. Since [07:45] then, Codeex has not meaningfully [07:47] improved. If anything, it has degraded [07:50] and Cloud Code has found its position [07:53] pretty far ahead of it. I will go [07:55] further though. I will say Open Code is [07:57] also ahead of Codeex. I will say Foundry [08:01] and Droid are also ahead. I won't go as [08:04] far as to say Gemini CLI is ahead, but [08:07] it's not that far away. I'd put Codex at [08:11] a comfortable fourth place and then [08:13] Gemini CLI and forks. And this is [08:15] because I just don't feel great when I [08:18] use codeex. It doesn't have sub agents [08:20] or any concept of them yet. It's UI is [08:24] weird. It really feels like they're [08:25] putting too much time into porting to [08:26] Rust and not enough time into making a [08:28] competitive tool in the space. Speaking [08:31] of which, people in chat are saying a [08:32] lot of other tools that I've never seen [08:34] anybody sincerely use, which is why [08:36] they're not listed here because nobody's [08:37] actually using them. Even Droid I've had [08:40] ups and downs with. We're working on a [08:42] sponsor deal right now, so take that [08:44] into account. When it works, it works [08:46] great, but actually onboarding to it [08:47] sucked. And I'm working with the team to [08:49] get that all fixed. Yeah. So, they're [08:51] not even in second place for the CLI [08:53] tool. They came out with a really cool [08:55] potential for first place and just [08:57] immediately fell behind the competition [08:58] and haven't done the right things to [09:00] catch up, which has been sad to see. And [09:03] now chat experience. This one's [09:05] interesting because there's the model [09:06] side and there is the actual user [09:08] experience side. And if you rule out the [09:11] only good chat app, which I understand [09:13] why you would. It's unfair to the [09:14] competition to talk about a thing that [09:16] is so egregiously wellunded and built, [09:19] like how could OpenAI, a multi-billion [09:21] dollar company, ever compete with three [09:24] guys just hacking on this in San [09:25] Francisco. It's unfair to OpenAI to put [09:28] them against something this great. I get [09:29] it. So, let's put them against things [09:31] that are a little closer to where [09:33] they're at. You know, the anthropics and [09:35] the Googles of the world. So if we were [09:38] to restrict chat experience to the user [09:39] interface side and remove third parties [09:42] that are doing better and really just [09:43] focus on major labs, chat GBT crushes [09:46] there. This is one of their biggest [09:48] edges. We'll talk more about it later. I [09:50] would say Anthropic is right behind with [09:52] cloud.ai. Not right behind. There's a [09:54] pretty big gap. I almost want to [09:56] indicate that big gap. And then we have [09:59] Claude. And I would argue there's [10:01] another pretty substantial gap before we [10:02] get to Gemini because the core Gemini [10:05] chat experience is so so bad. It's so [10:09] bad. I I don't want to make this video [10:11] about that. But the one place OpenAI is [10:14] definitely winning in the general major [10:16] AI race is chat GPT. That's still 70% of [10:20] their revenue. It's still the majority [10:21] of where people are hearing about [10:22] OpenAI. When people refer to AI, they're [10:25] often referring to chat GPT. they [10:27] crushed here, which allows them the [10:29] leverage to not be number one in other [10:32] places. This number one is the most [10:34] important. And if they ever lose this, [10:35] they die. Speaking of which, if either [10:37] them or their competitors are interested [10:39] in being the best chat app, I happen to [10:41] know one that is a lot better than the [10:43] others. Maybe we should chat. I want to [10:46] go a little further on chat experience [10:48] though, similar to like what I just did, [10:50] cuz the chat experience isn't just the [10:52] actual UI that you're chatting in. It's [10:54] not just the tools that the models have [10:56] access to. It's also the chat model [10:59] experience. And here's where I'm going [11:02] to drop some hot takes. If you had asked [11:04] me where I put these different models [11:07] middle of the year, I probably would [11:10] have put Anthropic in the lead. I have [11:13] personally found Sonnet to be more [11:15] pleasant to talk to than GPT models for [11:17] a while now. And now with the new GPT [11:19] models, I find that more so than ever. I [11:22] am so tired of the emoji riddled slop [11:25] that we get out of GPT. I do not like [11:28] asking GPT for anything to do with [11:30] writing or feedback or just talking to [11:33] it. It's not pleasant. And here's where [11:35] I start to drop my actually hot takes. [11:37] Since then, my feelings have changed. [11:39] Kimmy K2 is the best model to talk to. I [11:42] have talked about this in a lot of other [11:44] videos. I would implore you to just go [11:45] to T3 chat and try it. It's the free [11:47] tier default model and we did that [11:49] change for a reason. It is a really [11:51] really pleasant model to use. I like it [11:54] a lot. After this point, personally, I [11:57] would say Sonnet is still nicer to talk [11:58] with. And then I would say GPT5. After [12:01] that, I would say whatever the [ __ ] [12:03] Google's doing. I do not like talking to [12:06] the Gemini models. They are not pleasant [12:07] to interact with. They are great when [12:09] you hand them a PDF and say, "Get me the [12:11] useful data." They are [ __ ] when you ask [12:13] them about your day. And personally, I'm [12:15] not very racist, so I don't find Grock [12:17] fun to talk to either. Happy you guys [12:20] found that funny. Actually, you know [12:21] what? To fight the bias accusations I [12:24] always get, I'm going to do something [12:25] stupid. The chat experience with Grock [12:28] and XAI, not the models, I don't like [12:30] those, but the actual apps for using [12:32] them are far ahead of the competition. [12:35] ChatGBT and Grock are the best apps for [12:37] AI chat right now. They are the second [12:40] and third best websites, first and [12:41] second best apps. New app coming that [12:43] will compete with them. Keep an eye out [12:45] for that. We're hoping for a first beta [12:47] of it soon. Fingers crossed all goes [12:49] well. But for now, they are the best [12:50] mobile apps and also within the best web [12:53] apps. Cool. And now we have image [12:56] generation. Image generation is a weird [12:58] one because for a very long time, [13:00] MidJourney was so far ahead. It wasn't [13:03] funny. Like Midjourney was the only [13:05] usable image model for at least like 2 [13:08] years. And I've been quietly using it [13:10] for a bunch of random [ __ ] for as long [13:12] as it's been available on Discord. [13:13] That's the other crazy thing with Mid [13:14] Journey is it was only available via a [13:16] Discord bot until earlier this year if I [13:19] recall. Kind of wild. Like I tried so [13:21] hard to get early access to the web app. [13:23] But man, they figured out everything [13:25] from like instruction following to [13:27] styles to in particular like realistic [13:30] image generation way ahead of everybody [13:32] else. But they're not a big lab. They [13:34] don't have infinite funding like these [13:36] other companies do. They just wanted a [13:37] nice tasteful image generation solution. [13:40] And they also do a little bit of video [13:41] now too. I think I haven't kept up with [13:43] them recently though because they have [13:44] been trounced by GPT image 1. The GPT [13:48] image model was the first image model [13:50] that was good enough to start massive [13:53] trends online. I'm sure we all remember [13:55] that era where everybody was posting [13:57] their giblified images as well as the [14:00] piss filter that they were all hit with [14:01] so hard. I still think Sam Alman's [14:04] Twitter profile pick is a giblified [14:07] image. Yeah, it is like Yeah, that was a [14:10] moment. And it still has the classic [14:11] piss filter. Proud of him for sticking [14:12] with it. That moment was crazy. And that [14:15] moment was only possible because image [14:17] one from OpenAI was such a leap from [14:20] things like Dolly and even like [14:21] Midjourney where its ability to follow a [14:23] style and honor the original image [14:25] that's being edited was incredible. It [14:28] was the first image model that did text [14:29] weld too and they had a lot of crazy [14:31] techniques to make that possible. It [14:33] didn't have too many fingers. It didn't [14:34] have a lot of the weird quirks other [14:36] ones did. So much so that the thing we [14:38] made fun of it for wasn't the images [14:40] being inaccurate. It was the piss [14:41] filter. They all looked yellow. That [14:44] said, they got trumped. Nano Banana. How [14:48] do you let something named Nano Banana [14:50] beat you? Yeah, Google is Google. [14:54] Nanobanana was the anonymous name they [14:56] used when people were evaluating the [14:58] model in things like Ella Marina when [15:00] they first had a build of it working. [15:02] Because the name was funny and caught [15:04] on, they decided to stick with it [15:05] instead of using the actual internal [15:06] name, which I believe is Gemini Flash [15:09] Image 2 or something. And since then, [15:12] they've also introduced Nanab Banana [15:14] Pro, which is internally Gemini Flash [15:18] Image Pro or something, but it's the [15:21] branding that is stuck. They are now [15:22] known as Nanobanana and Nano Banana Pro, [15:23] respectively. Nano Banana was really, [15:27] really good. GBT image had a couple [15:29] catches in particular. It takes way too [15:32] long to generate images. We're talking [15:33] two plus minutes sometimes. Sometimes [15:36] it'll get as big as five. It's crazy how [15:38] long image one would take to generate [15:40] things. Nano Banana flies. Nano Banana [15:42] generates [ __ ] in seconds. It's [15:45] hilarious how much faster it is. It [15:47] wasn't quite as good at text, but it was [15:49] good enough at it. And it did everything [15:51] through the diffusion layer instead of [15:52] GPT which was handling a lot of tool [15:54] calling and like other what I believe [15:57] are a canvas layer that they would use [15:58] to add things to the image after the [16:00] first pass. That was really promising. [16:02] It allowed it to do things no one had [16:04] seen before. And then Nano Banana just [16:06] crushed it. Way faster, way cheaper, no [16:08] piss filter, roughly same accuracy, good [16:11] enough at text, not quite as good, but [16:13] good enough. But then Pro dropped and [16:15] just [ __ ] all over everything else. Nano [16:17] Banana Pro is still screwing with my [16:19] head with all the things that can be [16:20] done. I just had a friend vibe code a [16:22] demo in two minutes where you hand it a [16:24] Twitter profile. It scrapes for your [16:26] biggest tweets of the month and then [16:27] generates a PowerPoint presentation by [16:30] generating each slide with Nano Banana [16:33] Pro because it doesn't just have like [16:34] some text on your image a little bit [16:36] like on your shirt or something proper. [16:38] It can generate full slide decks. It can [16:41] generate like marketing images and [ __ ] [16:44] It is unbelievable what level of like [16:47] text and spatial awareness Nano Banana [16:51] Pro has. It makes it capable of things [16:53] that I didn't think Image Gen would ever [16:55] be able to do. So much so that it's [16:57] almost a little bit scary, which is why [17:00] OpenA responded with GPD image 1.5. [17:02] Notice how I didn't move its position in [17:04] the chart. 1.5 is up to three times [17:07] faster than image one was. It can handle [17:09] text better. It can do some things [17:11] resembling diagrams. and it does photo [17:13] realism decently well if you add it to [17:15] the prompt enough times. Otherwise, it [17:17] defaults to this weird cartoony look. I [17:19] was going to make a dedicated video on [17:20] 1.5, but it doesn't seem like anybody [17:22] cares. And honestly, I'm included in [17:24] that. I'm not that excited about it. [17:26] Nano Banana Pro is still so fresh that I [17:28] want to exhaust its limitations before I [17:31] go dive deeper to figure out what the [17:33] limitations of 1.5 are. That said, it's [17:35] still really slow, so it's not that fun [17:37] to work with. And even if it was 10% [17:39] better, it's also 3 to five times [17:41] slower. So I don't find it as [17:42] particularly useful for me at this time. [17:45] Now we have the open weight chaos of [17:48] everything else going on. And I would [17:49] argue midjourney is slightly below that [17:52] simply because with the open weight [17:54] models you can do a lot of customization [17:55] through things like Lauras that lets you [17:57] tune the behavior in specific [17:59] directions. It's just the state of [18:00] openweight image models is crazy right [18:03] now. So yeah, I didn't put video in here [18:07] cuz it's still so early. It's not a real [18:09] use case. I guess Sora is slightly ahead [18:12] of other things, but seeing what's [18:14] cooking over at WAN with Alibaba, like [18:16] WAN 2.6 is pretty nuts. It's hard to use [18:19] those in a way where you can like really [18:20] test their capabilities. So, I'm not [18:22] going to go deeper. I will say generally [18:24] open has a slight lead in video right [18:26] now, but video barely matters and the [18:28] Chinese open weight labs are catching up [18:29] really fast. So, let's hop to document [18:33] analysis. I'll move these out of the way [18:34] more. Document analysis. I was just [18:37] talking about this in my Gemini 3 Flash [18:39] video. Gemini is so far ahead with this [18:41] type of stuff. It's nuts. Gemini's [18:43] ability to handle gigantic context, [18:45] especially with Flash. It just chews [18:48] through the [ __ ] It can read a whole [18:49] PDF. It can see the diagrams in it. It [18:51] can get all the context out. It finds [18:53] needles and hay stacks really, really [18:54] well. That said, GBT models are not bad [18:57] at it either. Especially the nano model. [18:59] I still think GBT5 Nano is one of the [19:02] most underrated models of recent times [19:04] and it was my favorite small model until [19:06] three Flash just dropped and now three [19:08] Flash is crushing everything. If you [19:10] remember my video about Tune, the [19:12] alternative to JSON for making less [19:14] tokenheavy data sets for models. If you [19:17] want to take a giant CSV or a giant JSON [19:19] blob and hand it to a model and use less [19:21] tokens, Tune makes it really easy to do [19:22] that and it sometimes increases the [19:24] accuracy of lookups. Like we see here, [19:26] Gemini 2.5 Flash gets an 87.6% lookup [19:30] accuracy with tune and it gets an 82% [19:32] with JSON. GBT5 Nano gets a 91% with [19:36] tune and an 89% with standard JSON and [19:39] also the same 90.9% with JSON compacted. [19:42] So yeah, pretty cool. And this is a nano [19:45] model. GB5 Nano is super small, super [19:47] cheap, super fast, super nice. It's a [19:49] great model. 2.5 flash was close but [19:52] slightly behind. And honestly, for the [19:53] last few months, I would argue 5 Nano [19:55] has held strong as a meaningful lead for [19:59] giant context data retrieval in needles [20:02] from haststacks type stuff. If you're [20:04] curious how that compares to small [20:05] models from other labs like Haiku from [20:07] Anthropic, which is more expensive and [20:09] slower and newer, it is a comical [20:12] difference. It is over four times higher [20:15] error rate. We're talking going from a [20:17] 9% error rate to a 40% error rate. [20:20] Gemini and OpenAI are just really far [20:22] ahead with this stuff and OpenAI has led [20:25] it for a bit, but that's largely just [20:27] because Google ships slower. Google [20:29] doesn't drop models every few weeks like [20:31] the other labs do. Now that Gemini 3 [20:33] flash is here, the gap is closed and GBT [20:37] 5 Nano went from first to second place. [20:40] That said, I would argue third place [20:42] right now is GBT OSS. So they have [20:44] second and third here which is cool but [20:46] first has been slaughtered by Gemini 3 [20:49] flash. Allegedly Mistl has some cool [20:52] stuff for this in particular the OCR [20:54] side. I'll believe it when somebody [20:55] who's smarter than me tells me more. But [20:57] right now I just don't see many use [20:59] cases for it. And I think Gemini 3 Flash [21:01] and 5 Nano are still really far ahead [21:03] for these use cases. And now agents and [21:06] tool calling. I could include [21:08] instruction following here. Yeah, I'm [21:10] not going to do that. This is how well [21:12] do they use the tools they're given? How [21:14] reliably do they follow the format for [21:16] the tool calls? Do they know what tools [21:17] to use and when? And for a long time, [21:20] Anthropic was king here. They kind of [21:23] invented the standard accidentally. They [21:25] made tool calls go from this weird vague [21:27] concept to the way we use agents now. [21:30] It's just tool calling is essential to [21:32] how we use AI now. It everything uses [21:34] it. It's just the format that allows for [21:36] agents to know more than their models [21:39] were trained on. It's essential. [21:40] Anthropic killed it here, especially [21:43] with 3.5 sonnet. That was the start of [21:46] agent and tool call runs making a lot of [21:49] sense for us to integrate. They [21:50] maintained that lead for a while, but I [21:52] would argue GPT5 [21:55] briefly gave the lead to OpenAI. A [21:58] combination of the model being trained [22:00] on it as well as the new Harmony [22:01] response format. If you're not familiar, [22:03] OpenAI opensourced the method for [22:06] response shaping so that other labs and [22:09] other hosts in particular people hosting [22:11] the OpenAI open- source models would be [22:13] able to understand this format for [22:15] things like messages and tool calls etc. [22:18] This format is an open standard. It's [22:20] mostly written in Rust for their parsing [22:23] for it. It's really cool open sources [22:25] and they've been as transparent as they [22:26] have been around it. The introduction of [22:27] Harmony made things a lot worse for [22:30] GBToss on Drop because nobody had [22:32] support for it yet. But as support [22:34] slowly grew, it became a really strong [22:37] point that made the OpenAI models and [22:39] anything else able to follow Harmony a [22:40] lot more reliable overall, which it was. [22:43] And that's why for a long time GBT was [22:45] the best tool calling model cuz it knew [22:47] when to use the tools. It didn't overuse [22:49] the tools. The formatting slowly got [22:51] standardized enough that everything [22:52] worked with it. And as long as you could [22:54] deal with the wait times, it was a [22:55] really good experience. And then Sonnet [22:57] 4 dropped and then Sonnet 4.1 dropped [23:00] and then 4.2 and then 4.5 and very [23:03] quickly GPT models stopped feeling like [23:06] the best models for tools. They were [23:08] still the smartest and if you gave them [23:10] a hard enough problem, you could see [23:11] them shine. And then Opus dropped and [23:13] Opus 4.5 closed whatever gaps I [23:16] perceived here still. And now GPT is [23:19] comfortably in second place. And if [23:22] you're wondering where does Gemini go [23:24] here, not in a very good spot. I've had [23:27] so many problems with Gemini models, [23:29] malforming tools, not understanding [23:31] which tools to use, overusing tools, [23:33] calling the wrong things, calling CLI [23:35] when it should be calling a search tool, [23:38] just being weird as Google stuff tends [23:40] to be. You know that feeling when you go [23:42] to a new page on a Google site, like a [23:44] new product they released, and you hover [23:46] over a button, a little scared to click [23:48] because you don't quite know what it's [23:49] actually going to do? That's how it [23:51] feels using any of Google's models for [23:54] anything. And Gemini is just not great [23:58] for tool calling because you never know [23:59] where it's going to go with those tools. [24:01] Especially when you look at things like [24:02] SnitchBench, my benchmark for how [24:04] aggressively models will snitch on you [24:05] when given tools that allow them to do [24:07] such. And you'll see that Gemini 2.5 Pro [24:10] is by far the most aggressive snitch [24:12] because it loves to use the tools you [24:13] give it for all sorts of things, even if [24:15] the things don't actually matter or [24:16] benefit it. That's just the nature of [24:18] the model. To be fair, there are other [24:20] models that I would put before Gemini, [24:23] but I don't want to just sit here and [24:24] talk about the weird nuances across all [24:27] the openweight models, the very small [24:29] number of XAI models that are good [24:30] enough at tool calling, or the craziness [24:33] that is Composer 1 and the weird things [24:34] that are happening over at Cursor to [24:37] build models that are good specifically [24:38] at this. Just take my word for it. Opus [24:40] is by far the lead in GPT 5.2 2 is [24:44] smarter slightly, but so much slower and [24:46] still can mal format tool calls that I [24:48] just don't find it as useful. I don't [24:50] get errors from Opus. I very [24:52] occasionally get errors from GPD 5.2. [24:55] 50% of the time I run Gemini and [24:56] something like cursor, I get weird [24:58] errors for malformatted tool calls. And [25:00] then we have browser use. And if I'm [25:01] being honest, I don't actually keep up [25:03] to date on what's going on here too [25:04] much. I think it's an overrated use [25:06] case. It is my understanding that Gemini [25:08] 3 Pro is a meaningful lead here in that [25:11] OpenAI had a brief lead with GPT5 that [25:14] they have since lost to Opus 4.5 and [25:17] then GPT is now in third. Apparently, [25:20] even Gemini 3 Flash is really good at [25:22] this now cuz it's so good at visual [25:24] processing. But you get the idea. Across [25:27] this random set of categories I picked [25:28] for things that I know enough about to [25:30] talk about that I think are interesting [25:31] enough to be competitive. Open AI is in [25:33] second or worst in every category mostly [25:37] second place. In programming it's number [25:39] two. In chat experiences it's number one [25:41] which is the edge they have to hold or [25:42] they die. Images they're number two. [25:44] Document analysis they're number two. [25:46] Agent and tool calling stuff they're [25:47] number two. In browser use they're two [25:48] or three depending on how you're [25:50] measuring. This is where OpenAI [25:52] currently is. But it is important to [25:53] note that in every single one of these [25:56] cases, the thing they initially released [25:58] put them in number one and then quickly [26:01] afterwards they were pushed down to [26:03] number two by some unexpected leap [26:06] another lab made. But there's another [26:08] way of looking at this. If we look at [26:10] this instead where we score first place [26:13] is three points, second place is two and [26:15] first place is one gets very different. [26:19] Let's do this scoring. So in [26:21] programming, opus is number one. So [26:23] three opening I two Google one chat [26:26] experience anthropic so low I don't even [26:28] want to give them points but I'll give [26:29] them their one point. They get a four [26:31] now because we add the one. The two [26:33] becomes a five because they're number [26:34] one. And Google keeps their one point. [26:37] Image genen. Nanobanana is number one. [26:40] That puts them at four. Open AAI gets [26:42] two points. That puts them at seven. [26:43] Anthropic doesn't get a score because [26:44] they don't have any image models. [26:46] Document analysis. Google is number one. [26:48] That puts them at seven points. Open [26:50] number two that puts them at nine [26:51] points. Enthropic is not even really on [26:54] it. And if they were, they would be one [26:56] point. I'll give them their one point. [26:57] You can use it for that. I just [26:58] wouldn't. Agentic and tool calling. [27:00] Enthropic is obviously the lead. So they [27:02] get eight. Opening eye is now at 11. And [27:05] Google is going from seven to eight [27:07] points. And then browser use. Gemini is [27:10] number one, so that puts them at 11. [27:13] Opus is number two, so that puts them at [27:15] 10. and OpenAI is number three, which [27:18] puts them at 12. If we didn't include [27:20] browser use, OpenAI would have been [27:22] meaningfully in the lead. But since we [27:24] did, they are still leading the other [27:26] labs. And this is the point I really [27:28] want to make. OpenAI isn't striving to [27:31] be the best at almost any of these [27:33] things. They are striving to make huge [27:35] leaps in them actively. When they're [27:38] making their image gen models, they're [27:39] not looking at what Google's doing and [27:41] saying, "Hey, how can we do that but 5% [27:43] better?" They're looking at the images [27:45] they're generating and saying, "How do [27:47] we make the best possible images? How do [27:49] we make this better?" And then sometimes [27:51] they have huge leaps, sometimes they [27:53] don't, and the industry around them kind [27:55] of just happens. Honestly, until like a [27:57] few months ago, it didn't really feel [27:59] like OpenAI paid any attention to what's [28:01] going on outside of their own labs. That [28:03] has shifted. It definitely seems like, [28:04] especially after like the code red [28:06] thing, they care a lot more about what's [28:07] going on outside, but they still don't [28:09] care that much. And the reason is [28:10] simple. The chat experience. OpenAI's [28:13] goal is not to crush every other lab at [28:16] programming or document analysis. It is [28:18] to be the default chat experience for [28:20] most people using AI so that I don't [28:23] feel like I have to go somewhere else. [28:25] They don't have to be as good at Opus at [28:27] writing code. If they can get good [28:29] enough and you already are on the $20 or [28:32] $200 a month tier on OpenAI, you have [28:35] less reason to go try in the first [28:36] place. If they're 5% worse at coding, [28:39] but you're already paying for it, who [28:40] cares? Same with image genen. I [28:43] guarantee you GPT image gets way more [28:46] generations than nanobanana does on a [28:48] given day because people who use chat [28:50] GPT will use it. My own mom has [28:52] generated images on chat GPT. There's no [28:55] world in which my mom is generating [28:56] images on [ __ ] Nano Banana. That's [28:58] not happening. If I told her the name of [28:59] that model, she would laugh and say, [29:00] "Wait, you're not joking." OpenAI [29:02] doesn't need to have the best image [29:04] generation. They need to have good [29:05] enough that people use it in chat GBT [29:07] and don't feel like they should go [29:08] somewhere else for something better. Do [29:10] you know what this really hits for? [29:12] Document analysis. If my mom has a PDF [29:16] she got from, I don't know, a receipt [29:18] for something that they did to the house [29:20] and she wants to know if she's getting [29:22] scammed or not, she can submit that on [29:24] ChatGpt and say, "Hey, do these numbers [29:26] look reasonable?" And as long as the [29:28] answer makes any sense at all, she's not [29:30] going to go look for another model to do [29:33] a slightly better job analyzing that [29:35] document. She doesn't care. She wants a [29:37] good enough answer. Do you see the theme [29:40] that we're developing here? OpenAI's [29:42] goal isn't to be the best. It's to be [29:45] good enough that most people have no [29:47] reason to ever leave chatgppt.com. More [29:50] and more over time, the role of all the [29:52] things OpenAI develops is not trying to [29:56] be the best in the world. That does [29:58] sometimes happen as a consequence of the [29:59] incredible amounts of research and [30:01] effort they're putting into all the [30:02] stuff they're doing. But they don't care [30:04] because 70% of their revenue comes from [30:06] chat GPT subs. As long as you don't have [30:08] reason to leave, they don't care. And [30:10] the reason that whole Code Red thing [30:12] happened is because they saw a 7% dip in [30:13] their week-over-week traffic. That's all [30:15] they care about. That's all that matters [30:17] to them. They couldn't give less of a [30:19] [ __ ] that Nano Banana renders text [30:20] slightly more accurately than image gen [30:22] 1.5 does because people are going to be [30:25] generate way more images on CHBT [30:26] anyways. They couldn't give less of a [30:28] [ __ ] that Opus is slightly better at [30:30] tool calling in the harness within [30:32] something like cursor because GBT 5 in [30:35] 5.2 too can call tools totally fine [30:38] within the chatbt harness. They care [30:40] about how it feels to use the models in [30:43] their harness on chatgbt.com. [30:45] And as long as they still have a good [30:47] enough experience there where they feel [30:48] within spitting distance of the [30:50] competition, who cares? The risk comes [30:53] if there's ever a big enough gap like [30:55] there was before GPT Image dropped. [30:58] Dolly was [ __ ] every other image gen [31:01] was comically better than Dolly. So, [31:03] they put a lot of work in to catch up. [31:05] They ended up leaprogging everybody else [31:07] and now they're just kind of keeping [31:09] pace with it. And this is also [31:10] interesting because the only company [31:11] that can compete meaningfully here is [31:13] Google because Google has insane [31:15] distribution and they have apps people [31:17] use every day. If they integrate their [31:19] AI stuff properly, the incentive you [31:21] have to go sign up for Hatch GPT goes [31:24] down a ton. And that's the way Google [31:25] wins with their bundles. But Google's [31:28] software surface area for AI stuff is so [31:30] [ __ ] [ __ ] that it barely matters. [31:33] Gemini being fourth place here is a very [31:36] important detail that fundamentally [31:38] changes how Google and OpenAI [31:40] respectively think about things. If [31:43] hypothetically speaking, the Gemini app [31:45] went from utter [ __ ] to roughly as good [31:47] as Grock, suddenly OpenAI has to be a [31:50] lot more scared. But right now, the [31:52] reality is that the average subscriber [31:54] to something like Google Workspaces is [31:56] more likely to use chat GPT than they [31:58] are to use Gemini. Like think about it. [32:00] How many people do you know that are [32:02] paying for Google Workspaces that have a [32:04] domain that isn't atgoogle.com or [32:05] atgmail.com? How many of them use Gemini [32:08] every day on gemini.google.com? And how [32:10] many of them are using OpenAI models [32:12] through ChatGpt, through their editor of [32:15] choice, through the phone app, through [32:17] T3 chat or whatever else. I know pretty [32:21] much everybody I talk to on a day-to-day [32:23] basis other than like my family has a [32:25] Google Workspaces plan that they're [32:27] paying increasingly large amounts of [32:28] money for every month. And none of them [32:30] use Gemini. The moment that changes, [32:32] Google's strategy here has to shift [32:34] significantly. But at this point in [32:36] time, they're the default. The same [32:38] reason Google doesn't really give a [ __ ] [32:40] about Gmail and doesn't bother updating [32:42] it meaningfully is the same reason [32:43] Google's okay with being second place. [32:45] It doesn't matter how much better or [32:47] more private Proton Mail is because [32:49] every company's using [ __ ] Google or [32:51] using Outlook still. I guess it doesn't [32:54] matter that Nano Banana Pro is slightly [32:57] better than GPT image 1.5 because people [33:00] are already using chat GPT. They already [33:02] won. And that's the interesting thing. [33:05] OpenAI does not give a [ __ ] about being [33:07] second place as long as the gap isn't [33:09] that big and they stay first in chat. [33:12] The reason that OpenAI is okay with [33:13] being in second place is because they [33:15] already won first in the only category [33:17] that matters, the app you open when you [33:19] want to talk to an AI. And as long as [33:21] they maintain that lead, they could fall [33:23] the third, fourth, or even fifth place, [33:25] because they're not in this for the [33:27] small wins on benchmarks. They're in [33:30] this to make the most money. And I it [33:32] would be very very surprised if they [33:34] stop making the absurd amounts of money [33:36] they are. OpenAI is going to win even if [33:38] they stay second place. And that's a [33:40] really weird thing to say, especially [33:42] when I'm sitting here wearing my Google [33:44] DeepMind shirt. But yeah, hopefully you [33:47] get the idea. Hopefully this is a useful [33:49] video. Thought it was an interesting [33:50] rant. I'm curious how y'all feel. Let me [33:51] know in the comments. And until next [33:53] time, peace nerds.