In a fresh development, the most simple explanation I’ve heard for how an AI remembers things goes like this: Imagine a long, narrow table, and then imagine that you begin placing dominos on one end, slowly sliding them toward the opposite end as you add more.
According to the latest update, if the conversation goes long enough, you’ll start hearing the clatter of the oldest dominos falling to the floor, as they’re nudged off the table by the newer dominos. So, you keep chatting and chatting, placing more dominos onto the table while the very first dominos you added slide closer to the other end.
In a fresh development, today’s LLM models have limited context windows that can fill all the way up — and when that happens, something has to go, namely the very beginning of your chat. That’s a very rough description of how context windows work in AI.
According to the latest update, claude, for example, will occasionally pause a chat while it crams the context down to a more manageable size. Now, the ChatGPT, Claude, and Gemini apps all have methods for dealing with context windows that get too full, automatically “compacting” a conversation that’s gotten too long.
According to the latest update, but even with the best AI tools, you can’t ever tell precisely when they’ll decide to compact a lengthy chat, and while some AI apps will let you trigger a “compress this chat” session (Claude Code and Codex come to mind), mainstream AI applications like ChatGPT typically don’t.
As part of the ongoing story, that said, there is a way to manually compact a long AI conversation and then hand the distilled nugget to a fresh chat, essentially rebooting the thread (while saving precious AI tokens in the process).
In a fresh development, now, you could simply ask ChatGPT something like this: “Summarize our conversation so far into a tight 100-word paragraph,” but that risks losing important nuances, including critical forks in the road or even the original problem you’re trying to solve.
The report highlights that instead, try this prompt (which distills similar “compact this chat” prompts floating around the web):.
In a fresh development, include: what we’re trying to accomplish, key decisions we’ve made, anything important you’d get wrong by guessing, and the very next step. Keep it short enough to read in under a minute. Create a handoff summary I can paste into a fresh chat.
In a fresh development, this prompt should yield a brief overview of a long conversation that nonetheless carries over the most important nuggets of information, perfect for teeing up a fresh fresh chat.
The report highlights that i tried this prompt for a lengthy discussion I was having with ChatGPT about porting over a classic Apple II role-playing system update to a Raspberry Pi 5 (yes, it’s another one of my nerdy projects).
Industry observers note that we’re exploring how to get the original Apple II Wizardry: Proving Grounds of the Mad Overlord running legally on a Raspberry Pi 5, likely through Apple II emulation, not by rebuilding the system update from scratch.
As part of the ongoing story, a faithful clone or remake would be possible in Godot/Pygame, but the immediate goal is the authentic original Apple II version. Key decisions so far: the Pi 5 is easily powerful enough; the main issue is not devices but legal access to the original Apple II disk/programs.
The report highlights that wizardry is still a live IP, and random .dsk/.woz disk images online are probably unauthorized. The cleanest legal options are either buying the modern licensed remake for gameplay, or buying an original physical Apple II copy and imaging the disks yourself. The latter is the most authentic but may require specialized devices such as Applesauce or KryoFlux, and old disks may be degraded or copy-protected. Important caveat: don’t assume “abandonware” downloads are legal.
As part of the ongoing story, if authentic emulation, research current availability/prices for original Apple II. Very next step: decide whether the goal is authentic Apple II emulation on the Pi or simply legal gameplay.
According to the latest update, wizardry disks and what devices/programs is needed to image them into a usable disk image.
According to the latest update, then I handed off the conversation to a fresh chat thread like so:.
Industry observers note that let’s continue this chat, here’s where we left off: [Paste in generated summary].
Industry observers note that give it a try the next time your ChatGPT conversation gets too long, and you’ll help prevent any key context “dominos” from slipping off the table.
As part of the ongoing story, his coverage of artificial intelligence interrogates the most recent LLMs, and how they can be used at work and at home to be best prepared for the AI revolution. “AI is going to change our lives sooner than we think,” Ben writes. “Our best way to adapt is by using it every day.” Ben has been a PCWorld author since 2014, and has covered everything from laptops to security cameras before launching PCWorld’s AI beat. Ben's articles have also appeared in PC Magazine, TIME, Wired, CNET, Men's Fitness, Mobile Magazine, and more. Ben holds a master's degree in English literature. Ben has been writing about consumer technology for more than 20 years, and now focuses his reporting on AI as it relates to the basic human experience.