Aaron Walter

building @vellum_ai, prev: CTO @thinkfiveable (1M MAU), cmpe/cs/math @uwmadison

[Aaron Walter @awlevin]

@paulg This is how it felt to be in EdTech lol. Makes it 100x easier when the going gets tough

[Aaron Walter @awlevin]

@MichaelHewitt23 @marinatrajk Yes, you can also sign into your own email. We have Gmail OAuth to make it 1 click login, or you can ask your assistant to securely setup any another custom inbox (personally I have a nonstandard address 😭)

[Aaron Walter @awlevin]

@nickventuri Lmfao

[Aaron Walter @awlevin]

Fable 5 is insane. 15 minutes and two prompts, it built me a basic game that i can play with my *actual* vellum assistant.

you could imagine playing jackbox games hosted by your AI butler, that knows you and your family/friends well enough to customize the game perfectly for you.

but the part i'm most impressed with: most models suck at handling agent harnesses. Fable fully got the memo here— not only figuring out how to technically connect everything, but actually sending the right information at the right times. i was genuinely delighted by the proactive messages my assistant sent me in game: "hang on, let me heal you" and "two enemies on you, hang tight"

crazy times ahead. everyone should be building primitives to make it easier to vibe code extensions to their products

[Aaron Walter @awlevin]

@grim_tonal me: 90 seconds
fable: 15 minutes

it literally took longer for me to record the video and post the tweet than it took me to steer the model

[Aaron Walter @awlevin]

@SamuelBeek The last 20% is always the hardest lol. Might be in this weird jagged intelligence era for a while

[Aaron Walter @awlevin]

@_0xaryan @interaction @Apple I’d add that there’s a ton of slop hitting the App Store. Maybe Apple will start holding a higher bar for iMessage AI apps

[Aaron Walter @awlevin]

@maria_rcks @opencode @alan__rice Cause he be locked in

[Aaron Walter @awlevin]

@jaeholeeeee @Tekeee Metamask, crypto wallet

[Aaron Walter @awlevin]

My favorite framing is that, if models are trained on a corpus of a million libraries, and you prompt it with 1 measly sentence, you’ll probably get a similar outcome as someone else who prompted it with only one measly sentence.

If instead you give it your own unique ideas, with your own unique voicing, and an abundance of it, you’re steering it so much more forcefully to go in a different direction than anyone else’s 1 line prompts.

Give it your raw voice and opinions, and lots of them. That will make the outcome much more yours and much less the AI’s. The AI is just helping you wordsmith and better articulate your own ideas at that point.

Hell, we could probably *measure slop* by comparing the human provided context against what ultimately gets produced. Are most of the ideas coming from the AI? Slop. Most were expressed from human input? Not slop.

Whew

[Ole Lehmann @itsolelehmann]

Leila and Alex Hormozi just wrote an internal memo to kill AI slop in their company.

Here's what they said:

———————

"I’m going to be direct: I am SO sick of reading AI slop. Especially in memos.

So I would like to explain how you can immediately tell whether a memo is written by AI or a human.

At the end, I will tell you what you can do instead to write your memos.

THE TOP 5 PHRASES THAT INSTANTLY TELL ME AI WROTE YOUR MEMO:

1. “Delve into” / “Unpack”

Let’s delve into the Q1 results and unpack what this means for our strategy.

Nobody says “delve” in real life. You’d say “look at” or “break down.”

2. “This signals that” / “This underscores”

“This signals a shift in customer behavior and underscores the need for a revised approach.”

That’s AI connecting two ideas for you because it has no actual opinion.

A human would say, “Customers are doing X, so we need to change Y.”

3. “Navigate the complexities of” / “In an ever-changing landscape”

“As we navigate the complexities of scaling in an ever-changing landscape...”

This says absolutely nothing. What complexities? What’s changing?

4. “Synergies” / “Leverage our learnings” / “Holistic approach”

“By leveraging our learnings, we can unlock synergies across teams and take a more holistic approach.”

This is corporate bologna. It sounds profound and means NOTHING.

A human would say: “Here’s what worked, here’s what didn’t, here’s what we’re changing.”

THE TOP 5 THINGS CHATGPT DOES THAT GIVE IT AWAY IMMEDIATELY:

1. The transition words. Every paragraph opens with “Moreover,” “Furthermore,” “That said,” “Additionally,” or “Importantly,” Read your draft out loud. If you wouldn’t say it, don’t write it.

2. Bold word: colon: explanation built points. Example: Clarity: Ensure your memo is clear and concise. Alignment: Make sure all stakeholders are on the same page. Execution: Focus on actionable next steps.

3. The corporate therapist voice. This is a powerful opportunity to lean into our strengths and foster a culture of accountability.

4. The neat little bow at the end. “Ultimately, the goal is to build a more resilient and agile organization.” If your conclusion could apply to any company on earth, it’s wasted words.

5. The “says everything, means nothing” BS. You read an entire paragraph, it sounds smart, and then realize you can’t summarize what it said.

That’s because AI is pattern-matching language, NOT THINKING.

SO HERE’S WHAT I WANT YOU TO DO INSTEAD:

1. Use voice dictation to write your first draft. Talk through your idea out loud, in your own words, with your OWN voice. That’s your raw material, and it already sounds human because it came from one.

2. Use AI to make it better, not to replace your thinking. Ask: “What’s missing?” “Where is my language unclear?” “Help me tighten this without changing my voice.” “Does my argument flow logically?”

3. Do NOT paste in a topic and ask AI to write the memo for you. Also urants to add: forced negation (“not this but that”), staccato repetitions, and excessive adverbs like “quietly underscores.” All AI slop.

TLDR: The bullets you sent to AI to write your memo are arguably more valuable than the memo. I’d rather get those than the slop."

[Aaron Walter @awlevin]

@V3_Willy @gregisenberg I get that slack is built for 300+ but I also dont quite see where it’s is that much worse for 3 person teams. Any specific areas come to mind where you think it should be lighter weight?

[Aaron Walter @awlevin]

@adisingh What do you see in the future almighty Adi?

[Aaron Walter @awlevin]

@atmoio In macro econ this is called “weak links”. One example is that we have >1,000,000x transistors and performance in computers over the past few decades, but did GDP/productivity grow proportionally as much? Not even close.

Great video on it here: https://t.co/s63NjHrjrY

[Aaron Walter @awlevin]

@hthieblot Miniclip

[Aaron Walter @awlevin]

@ForrestPKnight @thdxr I think they do. Every company writes software. Many teams invest in employees to help their internal software ops (like devops). Building something like Cursor or Claude code is just another example of making your internal SWEs more efficient w/ something you can sell externally

[Aaron Walter @awlevin]

@peteoxenham Please someone please get me off of this waitlist so I can do funny things with my Taco Bell ultramarathon training plan

[Aaron Walter @awlevin]

@compileandpush @b_nnett Go away you sloptropolis

[Aaron Walter @awlevin]

@zenorocha Funny that in both cases it’s the inverse that should be celebrated (less tokens & less lines of code are more desirable)

[Aaron Walter @awlevin]

This is default, have seen it especially prevalent among Anthropic models. Some OSS models are distillations of that so it would make sense for those to exhibit it too.

One guess is that many viral social media posts in the training data were controversial topics (which get more engagement). So maybe models learned to express both sides of ideas during RL.