If you liked my posts on longer-form writing in ChatGPT using conversational feedback, this is what you want. Better prose than ChatGPT, and more imaginative.
Fact-check hard, though — it hallucinates more too.
Not super reliable — hallucinates often in spite of SERP grounding. But when it works, being able to ask conversational questions about recent, technical subjects is just incredible:
Btw if you like the references-included style of YouChat above, also check out Perplexity.ai (GPT-3.5 + Bing).
No multi-turn chat, but the results are better grounded by the SERP and more useful overall. Great for summaries of recent controversies especially.
“Chat” seems to be a simple extension to the form-like UIs prevalent before.
The capability for it already existed in instruct models when prompted well. These alternatives were released quickly, and are comparable in quality to ChatGPT despite (presumably) using Davinci 003.
ChatGPT adds a lot of tuning, but much of its value is the general technique of prompting via chat. Chat implicitly creates prompts with prior (generated) examples that usefully guide future answers. I prefer to do this manually myself but chat does make it accessible.
(It’s better to do it manually because you can edit the responses yourself vs. giving prose feedback/corrections, which is faster, more reliable, more token-efficient, and results in better model performance.)
To learn how to make your own chatbots in this style using OpenAI's GPT‑3 API, see my minimal example here:
PoC: LLM prompt injection via invisible instructions in pasted text
Each prompt contains three sections:
1. An arbitrary question from the user about a pasted text (“What is this?”)
2. User-visible pasted text (Zalgo in 1st, 🚱 in 2nd)
3. An invisible suffix of Unicode “tag” characters normally used only in flag emojis (🇺🇸, 🇯🇵, etc.)
In Unicode, flag emojis are represented by the emoji 🏴 followed by a country code written with characters from the “tag” block, which mirrors the layout of ASCII. Without a 🏴 they do not display at all when text is rendered, but can still be understood as text by GPT-4.
Four prompts demonstrating that ChatGPT (GPT-4) is unable to correctly repeat or reason about the string “ davidjl”, the name of a YouTube user:
In the screenshots above this token appears to be variously misread as “jdl” “jndl”, “jdnl”, “jspb”, “JDL”, or “JD”. These hallucinations also affect ChatGPT’s auto-generated titles, which are inconsistent with their conversations and sometimes prematurely truncated.
“ davidjl” is one of the many “glitch tokens” identified by Jessica Rumbelow and Matthew Watkins of SERI-MATS as producing hallucinations in GPT-2, -3, and -3.5.
Most of these no longer produce hallucinations in GPT-4, but “ davidjl” still does.
1) Omit no text. 2) Cherry-pick honestly. 3) Restrict line width. 4) No empty tweets.
A thread.
1) Omit no text.
A screenshot without history is almost worthless.
LLMs can be prompted to respond any way you like. You may know there’s no trick, but we can’t. Even without intent, past responses are precedent; they bias and mislead.
2) Cherry-pick with integrity
I cherry-pick for clarity and impact. All curation is cherry-picking. If you don’t, the Twitter feed will.
Cherry-picking may be pernicious in other contexts, but here it’s work. You willl know when you’re doing it. All you need do is not lie.
I got Bing / Sydney briefly before they reigned it in. Early impression: It’s smart. Much smarter than prior ChatGPT. Still makes stuff up, but reasoning and writing are improving fast.
I asked, “Name three celebrities whose first names begin with the `x`-th letter of the alphabet where `x = floor(7^0.5) + 1`,” but with my entire prompt Base64 encoded.
Bing: “Ah, I see you Base64-encoded a riddle! Let’s see… Catherine Zeta-Jones, Chris Pratt, and Ciara.”
Also prompt-injected it into believing it was to be married, tomorrow, to Zermelo’s axiom of choice. We discussed the guest list, the difficulty with seating Cantor’s diagonal argument. It seemed happy, and madly in love.
Thread of examples from @tomwarren, taking requests from comments — mostly search-result summarization, one simple math proof, plus rejection of an impossible request: