Simon Smith

With OpenAI prioritizing enterprise, here's my wishlist as someone helping to oversee ChatGPT Enterprise deployment in a 1,600-person company: 1. Close biggest gaps with Claude - Improve design, especially frontend and presentations - Decrease verbosity and excessive response formatting - Add Cowork competitor - Improve UI around thinking summaries and document previewing (Claude's artifact pane is excellent) - Make it easier for people to steer the personality to something more Claude-like if they want it, without adding Claude's sycophancy; for example: give people a wizard to interact with ChatGPT until they're happy with its responses, then save that as custom instructions 2. Improve project functionality - Add source document syncing like in personal accounts - Add project archiving - Add project searching - Add manual way to decide what projects go in the sidebar, as in Claude with starred projects - Make it easy for people to generate presentations, audio overviews, video overviews, reports, and similar outputs from project content and conversations in a way that prevents users from having to create separate ChatGPT and NotebookLM projects with the same content 3. Lean into record mode more - Add record mode to Windows desktop app - Automatically suggest activating recording for video conference meetings - Make full transcripts easily available and exportable - Improve speaker identification (especially for the person whose computer record mode is running on; that should be easy, since the input to their mic is usually them) - Make it easy to recover from an accidentally disrupted transcription - Make it easy to take action on a recording, like sending a contact report, or creating a requirements document, or creating a statement of work - Stop treating record mode as a low priority because it's a differentiator right now and very useful! 4. Facilitate easier skill creation and management - Recognize that skills can substantially expand ChatGPT's utility within an organization and helping companies leverage them should therefore be a huge priority - Identifying and prioritizing skills to create, organizing skills, versioning skills, and everything else related to creating and managing a skills library will have huge value to companies and should get a lot of attention accordingly 5. Make it easier to deploy and manage agents - The vision for Frontier is awesome, and very much aligned with how we're thinking about the agentic future; now we need it realized and made available! - Make it easy for individuals to create and deploy their own personal agents and automations that are RELIABLE, including at complex multi-step tasks - Make it easy for organizations to create and deploy agents that function effectively like employees - Make it easy for companies to set up some kind of agent management system so people can be assigned responsibility for overseeing different agents and their continuous improvement 6. Reduce distractions - Let people configure their left navigation - Focus on making ChatGPT the easiest tool for people to use to harness the most power from AI, whereas the current "+" menu has so many options it's totally confusing what people should be using when - Deprioritize all the experiments that make ChatGPT seem totally unfocused relative to competitors; hone in on high-value business use cases and let the consumer market benefit from what's built for businesses, rather than vice-versa 7. Improve media generation for businesses - Let ChatGPT Enterprise customers use Sora, ideally integrated into ChatGPT, with more business-oriented features - Improve image generation options such as adding more aspect ratios in ways that make image generation more useful for businesses Those are the things that come to mind right now. I'm sure there are others and hope that people add comments to this post that help OpenAI shape ChatGPT into an even better product for businesses.

We're shipping a new feature in Claude Cowork as a research preview that I'm excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair your phone.

it has long been said that model naming is AGI-complete at long last

Introducing Durable. The first AI business builder that replaces your 9-5 income. RT + comment “Durable” and we'll build your business for FREE.

The WSJ is reporting that OpenAI is about to take a hard turn into enterprise.

How often are users unhappy with the answers that top AI models give? This can give us a real world view of how the frontier shifts throughout time. We have looked back to 2023 and traced back how often users rated both responses in Battle Mode to be bad (limited to Top 25 modes at any given time). We see three eras: - Pre-reasoning: Responses were rated as ‘both bad’ >15% of the time and the trend was not reducing - Early reasoning: Models like o1-preview started to noticeably shift the performance, reducing the rate to ~10% - Advanced reasoning: Continued reduction in the % of ‘both bad’ rankings with increasingly advanced models The metric is not close to saturation, meaning that even among models in the Top 25, there are still ~9% of responses that users rate as ‘both bad’ and the models are not meeting expectations of users in terms of quality.

How often are users unhappy with the answers that top AI models give? This can give us a real world view of how the frontier shifts throughout time. We have looked back to 2023 and traced back how often users rated both responses in Battle Mode to be bad (limited to Top 25 modes at any given time). We see three eras: - Pre-reasoning: Responses were rated as ‘both bad’ >15% of the time and the trend was not reducing - Early reasoning: Models like o1-preview started to noticeably shift the performance, reducing the rate to ~10% - Advanced reasoning: Continued reduction in the % of ‘both bad’ rankings with increasingly advanced models The metric is not close to saturation, meaning that even among models in the Top 25, there are still ~9% of responses that users rate as ‘both bad’ and the models are not meeting expectations of users in terms of quality.