Skip to content

Context Windows

A context window is like a short–term memory for AI. How can we make it long-term?

5 min read
a purple sea as seen through a window
🦉
Before Growth is a weekly newsletter about startups and their builders before product–market fit, by 3x founder and programmer Kamil Nicieja.

Current generative AI models, whether textual or graphical, still face significant challenges when it comes to maintaining context. While they can remember things within a single session, these models are quick to forget once the session is lost or a new one starts, necessitating the need to prompt them with relevant details all over again.

For instance, if I’m using GPT–4 to brainstorm marketing strategies for the launch of a new feature but fail to save the conversation or simply lose it, ChatGPT will not recall the context I’ve already shared. Details such as my business description or my target audience profiles will need to be re–provided when I approach OpenAI for help with launching another feature.

Custom instructions do offer some level of control over how ChatGPT responds, allowing you to set your preferences and have them remembered for future conversations. I’m not particularly fond of this feature, given that it requires you to prepare in advance by laying out as much context as possible up–front. This feels like a chore, detracting from the overall magic of the experience. It would be far more beneficial, I believe, if the model could summarize key points at the end of a conversation and store them as “core memories.” Ideally, it should learn from our conversations, but this seems unfeasible due to the immense context window needed for the model to retain everything we discuss, or the impracticality of retraining a personalized model for each user.

Another possible solution to the context problem is the implementation of integrations. For instance, an AI–driven sales manager could learn about your customers and product by linking it to your Google account and accessing your past emails. And integrations will be a huge game–changer in the world of large language models, opening up incredible opportunities for startups. When a company like Google launches an AI assistant, that assistant is limited to operating solely within Google. But imagine an independent assistant that could seamlessly integrate with other platforms. It could pull brand assets from your cloud storage, whip up UI mockups in Figma, and then send them over to an AI programmer to code the frontend.

💡
ChatGPT now has the capability to learn from your conversations. It retains information from previous chats, enabling it to deliver more pertinent responses. This is something I anticipated as necessary here and follow up on in Why a Newsletter.

This kind of cross–functionality is currently unachievable with the walled garden approach favored by large corporations, but it represents a significant potential avenue for disruption by generative AI. OpenAI knows this, too. You can tell by their recent announcement of the ChatGPT Enterprise offering. One notable feature under development is customization, which will allow companies to securely augment ChatGPT's knowledge base with their own data through integration with existing applications. I think this is a sound strategic direction.


Hey, where’s our weekly digest? Don’t worry—this week is a departure from our usual focus. Tech IPO season seems to finally be back, so we’re going to give it a quick look.

You may think that startups going public is irrelevant to our usual discussion on early–stage ventures. I beg to differ. Exits, after all, are a pivotal endgame for many founders and investors. It’s a symbiotic relationship: while the early–stage landscape can forecast the trajectory of public markets, those very public markets reciprocally impact the landscape for startups.

After a drought of 18 months since the last significant tech IPO, we suddenly have three. This is a huge change!

The companies spearheading this movement are:

  • Arm, which designs semiconductor chips found in 99% of premium smartphones.
  • Instacart, the on–demand grocery delivery service.
  • Klaviyo, a developer of marketing automation tools, complete with email, SMS, and push notification capabilities.

For anyone dissecting IPOs, it’s standard to scrutinize the S–1 documents that companies file with the U.S. Securities and Exchange Commission. These documents are exhaustive dossiers revealing a company’s financial health and strategic intent, effectively serving as their IPO announcement. They’re not the most riveting reads, so I’ve distilled their essence for you. (I’ll omit Arm from this analysis—semiconductors are beyond my wheelhouse, and the firm can hardly be called a startup.)

Instacart IPO

Here’s the S–1 if you want to read it. Instacart serves as a great example of what one might call a “zero–interest–rate startup.” Originating in 2012, the company embraced the capital–intensive model, reminiscent of Uber. Known for burning through investor funds, Instacart has amassed a staggering $2.9 billion over 19 funding rounds. However, in stark contrast to Uber, which went public in 2019 without profitability, Instacart adapted to market expectations. As of the first half of 2023, it boasts a positive cash flow, with $235 million in operating income and $242 million in FCF.


Related posts

Chaining Prompts

Chains go beyond a single LLM call and involve sequences of calls.

Chaining Prompts

Intelligence as an API

AI models are getting hard to beat when it comes to getting simple answers right at scale.

Intelligence as an API

Pairing with AI

Raise your floor with large language models.

Pairing with AI