Tom Spis

context is the new black

[Tom Spis @hey_tommy]

@NavyaVanigota @btibor91 That's what I first thought too until I re-tested Writing Blocks today — see my comment for deets.

[Tom Spis @hey_tommy]

Just tried the new Library integration in Writing Blocks, and it really does seem like the better replacement for Canvas.

Try this: make any Writing Block full-screen, and you'll see an Add to Library button, which saves it in the Library — and not merely as a copy, but a live single instantiation of that exact in-conversation doc! That means you can update the same doc in either location (Library or the conversation), and it's updated everywhere!

That is SO much better than Canvas — where a) there was no way to know/search for which conversations had a Canvas doc, and b) Canvas contents were NOT indexed, so ChatGPT search would have zero results if you tried searching for content that was only present inside a Canvas doc. These two things in combination were why I stopped using Canvas docs, despite their strengths as mutable context-referenceable at the time.

[Tom Spis @hey_tommy]

@bygregorr @btibor91 That's crazy-long in AI years. Also, Writing Blocks really does seem better.

[Tom Spis @hey_tommy]

@scaling01 Oh no. You poor soul. The injustice. 🙄

[Tom Spis @hey_tommy]

@nicdunz No, it's garbagemodal.

[Tom Spis @hey_tommy]

@ryanbrewer @nicdunz If "not personalized" means bypassing custom instructions, I'll take the 30s every time. I've got a category auto-classifier and router that's bootstrapped via custom instructions, and Fast Answers bypass it.

[Tom Spis @hey_tommy]

@nicdunz @ryanbrewer Absolutely.
I don't pay for a sub to get Google AI Overviews and to save OpenAI on inference costs.

[Tom Spis @hey_tommy]

Ah yes, the classic bundling + framing combo: time the end of the promo with the extra squeeze on the 5h Codex window, then call it “rebalancing to support more sessions throughout the week.”
@grok, you seeing support for a 5h quota squeeze beyond just the 2x promo ending?
@btibor91 @testingcatalog @koltregaskes — would appreciate a quick sanity check on the read.

[Tom Spis @hey_tommy]

@pigeon__s @Angaisb_ Nothing about "there can be no security through obscurity" says "therefore, let's just not have locks if they can be broken".
You are conflating and collapsing several heuristics at once, and vibe-applying them to an entirely different domain/area of concerns.

[Tom Spis @hey_tommy]

What you're referring to as "suspicious" is just basic game theory principles: a) do not make the first "bad-faith" move, b) if someone else makes a "bad-faith" move, you should also make one. This is not primarily a move to signal principles (even if that is likely also at least part of the calculus) — it is entirely contingent on others following the example, and playing in kind.

If another frontier lab decides to not follow the lead, the very reason for delaying the model release so as not to provide tools for adversarial attacks immediately disappears, as those tools are publicly available.

[Tom Spis @hey_tommy]

@Angaisb_ With all due respect, your reasoning is flawed here.
I strongly recommend you ask a model for epistemic/adversarial review, and genuinely engage in the exercise of challenging your claims to see whether they stand up to scrutiny.

[Tom Spis @hey_tommy]

If you think that this specific ChatGPT feature allows "editing [your] memory of events while keeping the memory of [your] reactions constant", then you can breathe a sigh of relief knowing that is not actually the case.

Individual messages and conversation branches in ChatGPT are immutable (i.e not modifiable). As soon as you edit one of your previous messages, it creates a completely new conversation branch, shown in the UI as variant indicators like '< 2/2 >'. So any response or back-and-forth that followed that edited message is now its own separate conversation branch, available only when you switch to back to the 1st message variant by clicking that '<' button — and once you do, the indicator will now show '< 1/2 >', and you'll see your previous unedited message, and any original back-and-forth that followed.

The same thing applies to regenerating response variants. As soon as you've regenerated one, that's now a new conversation branch. If you reply to that new response variant, the conversation will continue on that branch. If you then later decide to go back and switch the response variant to the first response version, and respond to that message instead, you've now gone back to a conversational fork in the road, and are now continuing the branch independently of the previous one.

TL;DR: unlike Google AI studio, which has a different purpose (a UI playground for model response testing/steering/etc.) and allows you to edit any previous messages (even model responses) without invalidating any future responses, ChatGPT, Gemini, and all major frontier lab product UIs do NOT allow the thing that you are worried about.

[Tom Spis @hey_tommy]

@btibor91 Phew! Thanks @btibor91 for following up on this, and any magic dust you may have sprinkled in the process 😉🙏

[Tom Spis @hey_tommy]

@lateinteraction yes, that is how branding works, Mr. Programming Not Proompting 🤓🙃

[Tom Spis @hey_tommy]

@koltregaskes "Refocus" and "Simplify" are keywords for "we're bleeding cash — let us start priming you for incoming enshittification by getting ahead of the backlash with spin to make you think you'll be getting more, when in fact, you'll be getting less."

[Tom Spis @hey_tommy]

https://t.co/N3ycoT3W9I

macOS and mobile apps not yet updated, so editing, regen, and model switch + regen still working there. I'm going to stop updating my macOS app, and turn off auto-updates for ChatGPT for my Android. I don't expect that to work for long.
Fury is an understatement.

[Tom Spis @hey_tommy]

Wow — it just got A LOT worse.
As of today, you can only edit and regenerate the last user and the last model response, respectively. Thought it was a bug, but no, it's the new normal, and OAI has already updated their support articles. I think this is the last straw for me.

[Tom Spis @hey_tommy]

On a more useful note, the macOS app still has the Switch menu (though possibly/probably not for long), and when using the iPadOS app (and probably the iOS app too), changing the model up top, and then regenerating does switch+regen (unlike when using the website).

[Tom Spis @hey_tommy]

@chadptg @btibor91 Yeah, this absolutely asinine.
If this is an intentional and permanent change, whoever is responsible, may they be blessed with four permanently deflated tires.

[Tom Spis @hey_tommy]

Seriously, more than a month has passed, and no one thought to mention that what they're calling " context learning" is literally the definition of in-context learning. They completely mischaracterize ICL as an exemplar-only learning method — which is just not how the overwhelming majority of the research community & industry think of ICL in the last 2-3 years.
The annoying thing is their actual work itself is pretty interesting — but unfortunately, it was completely overshadowed by me eye-rolling over their framing as I read it.