|
|
"I tried that. It didn't work." |
There’s a lot of trial-and-error going on with AI. If you tried something just 90 days ago, you can expect your experience trying the exact same thing to be different today. |
We’ve had four major model releases over the past three months or so: Anthropic replaced Claude Opus 4.7 with 4.8 and Fable 5, and OpenAI dropped GPT-5.5 and GPT-5.6 over the same period of time. |
A consequence of AI progressing this fast is that we must hold our impressions of workflows, features, and tactics loosely if we want to maintain an accurate map of what AI is actually capable of. Now. |
I call this my 6-month rule: in a world where capability expands exponentially, retest your assumptions every six months. But the model upgrades we’ve seen over the past couple of months are so significant that I think it’s worth revisiting things now. |
So today I want to surface a short list of interesting things beginners can do with AI that have notably improved over the last six months, including: |
Why not every experiment you try with AI will work (but could work tomorrow) What has improved significantly over the past six months (and how to try it) The smallest, most impactful habit you can keep to stay up-to-date
|
—@dharmesh |
|
|
Do This Every 6 Months |
When I wrote about the most common AI mistake I see people making back in April, my advice was to retry anything that doesn’t work today, six months from now. |
If anything, these last six months have solidified my belief in that idea. |
Competing AI companies are racing to crown their newest models as the most powerful ones on the market, only to be dethroned weeks later. Now we even have governments stepping in to review the strongest models before they’re released publicly. |
It doesn’t feel as if things will be slowing down anytime soon. And I understand why that might feel overwhelming to you. But it's also very exciting -- and a tremendous opportunity for anybody willing to filter the signal from the noise. |
Reminder: Sometimes the experiments work, and sometimes they don’t. The important part is that you keep experimenting. Even the most advanced AI users, like the ones on the bleeding edge I make an effort to keep up with, are trying new things on a regular basis with mixed results. |
Just like a bad book shouldn’t derail your reading habit, an underwhelming result shouldn’t derail your AI education. |
I’ve picked out a few mini-experiments for beginners to try below. Each of these experiments centers around a notable improvement AI has made over the past six months or so. |
Ready to try? |
|
|
Worth Trying: What’s Changed in AI? |
Every AI tactic listed below is one that has seen significant improvement over the last six months or so. I’ve included a prompt to copy-and-paste for each, so you can easily try each of the highlighted tactics yourself. |
This is meant to be a fun way for you to catch up on some improvements you might have missed -- there’s no pressure to achieve any specific outcomes. |
⭐ (If you only have time to do one thing this week, try #3) |
1. AI has made speech-to-text 10x better |
Speech-to-text spent decades forcing you to talk like a robot. If you wanted a question mark, you said "question mark" out loud. Voice mode, as recently as January, functioned a bit like a walkie-talkie: you talked, you waited for the delay to pass, it talked back. |
Today, the mic button in ChatGPT, Claude, or Gemini handles your pauses and punctuation automatically. The new GPT-Live voice mode goes further and can research something mid-conversation without skipping a beat. I went deep on this in the last simple.ai newsletter, and the short version is: the messy, unedited version of what's in your head is often the better prompt. It might be the most dramatic jump on this list. |
Try this inside ChatGPT, Claude, or Gemini: Tap the mic and ramble for two minutes about something real you're working on. Then say: "Okay, I'm done. Pull out what actually matters here, organize it, and tell me what to do next." |
|
|
2. Images generated with AI are more accurate now |
At the end of last year, I said AI image generators feel a lot like slot machines: type a prompt, cross your fingers, and hope the text isn't garbled. Now modern models from Google’s Gemini and OpenAI’s ChatGPT are a lot better. Not perfect, but a lot better. |
Infographics, labeled process diagrams, a one-page explainer you can hand to your team without apologizing for it -- there are a lot of newly unlocked options thanks to the improved image models. |
An underrated use case I discovered secondhand: screenshot your website or app and ask for a redesigned UI. I saw that clever designers were doing this as an alternative to the more traditional, token-heavy approach of connecting AI to the codebase directly. |
Try this prompt inside ChatGPT: "Create a clean image of a 5-step infographic explaining [your process], with a clear title and short labeled captions for each step." |
|
|
3. The AI desktop apps have taken big steps |
Early on, the ChatGPT and Claude desktop apps were just the website wrapped in a different window. That’s not the case anymore. |
I believe this year we are seeing the rise of the “Super App”. The idea is simple: A single app that houses multiple AI-powered capabilities. |
OpenAI recently made a big step in this direction by merging their chat (ChatGPT) and coding (Codex) interfaces into the same app. It makes a lot of sense to me. Most consumers don’t want to use a bunch of different apps to do different things and switch back and forth between them. |
Using a desktop application to interact with AI is a big step up. A browser tab only knows what you paste into it, while the desktop app can reach the actual files and folders sitting on your computer. |
I wrote about Codex and Claude Code separating beginners from power users. As these apps improve, it will become easier and easier for beginners to access the more powerful AI features that have (until recently) been living in the terminal. |
|
|
4. AI can now return the exact file type you need |
AI used to hand you the raw material and leave the last mile to you. Usually this meant it wrote the words, and you were left spending an afternoon pasting that info into slides. |
Now you can ask for the file itself. A real PowerPoint, a working spreadsheet, a formatted Word doc or PDF, ready to download and send. |
Try this prompt inside ChatGPT or Claude: "Turn these notes into a PowerPoint presentation I can download: [paste your bullet points]. Ten slides, one idea per slide, plus a title slide." |
|
|
|
|
5. Have AI tutor you, step by step |
Asking AI to explain something is useful, but it does the thinking for you. A week later, how much do you really remember? |
Perplexity's "Learn step by step" mode walks you through the topic with guided questions and hints, and adjusts to what you already know. It was limited to student accounts until February, when it opened to everyone, free. |
With the right prompt, you can create a similar environment in ChatGPT or Claude. But if you haven’t used Perplexity in a while, this is a good excuse to give it a shot. |
Try it: Visit Perplexity, click “Search” and choose "Learn step by step". This feature is free to use but you will need to create a free account to unlock access. |
|
|
|
|
Retry the Work that Matters Most |
The tactics listed above are just a starting point. I hope they inspire you to keep tinkering around with any new AI release that catches your attention. |
Watch for the experiments that matter most to you. Somewhere in your notes, or in a frustrated message you once sent a coworker, there’s a pile of AI experiments you tried and gave up on. Those are the things you want to retry every 3-6 months. |
If nothing comes to mind, write down every step of your workflow. Can any of those steps be improved or partially handled by AI today? |
Being diligent about retesting the workflows that actually save you time will pay you back more than any list of mine, because it's built from your needs. When you get overwhelmed, revert back to this habit. |
Maybe I can help you overcome your current sticking point? |
Hit reply on this email and tell me about your biggest struggle with AI. I read every reply (and occasionally reply to a few when I can) -- and I’d love to hear from you. |
—Dharmesh (@dharmesh) |
|
|
|
|
What'd you think of today's email?Click below to let me know. |
|
|
