Logan Abele

Helping you understand AI and implement it for your business
💼 Ex Finance | Scottsdale
✉️ [email protected]

[Logan Abele @LoganAbele]

25. Founder. Starting from 0 followers.

I’ve spent the last few months building ForgeDreams AI.

Some days I feel like we’re building something incredible. Other days I have absolutely no idea what I’m doing.

I’m going to document both.

Day 1.

[Logan Abele @LoganAbele]

https://t.co/rtB1NOLRR8

[Logan Abele @LoganAbele]

Stop benchmarking AI agents on reasoning. Meta's Muse Wonder just showed the test that matters: does it remember what you told it yesterday? Run that test on whatever you're building before you ship.

[Logan Abele @LoganAbele]

Serious question: what's the one subscription you'd cancel tomorrow if a free AI clone existed?

Photoshop? Salesforce? Something boring and expensive?

The guy who cloned Adobe did it because they charged him a cancellation fee one too many times. Spite is an underrated product strategy.

[Logan Abele @LoganAbele]

One developer just rebuilt Adobe’s entire Creative Suite with Claude Opus 5.5.

Seven apps. Open source. Free forever. PhotoCraft already has 23,000 stars on GitHub.

Adobe spent decades building its moat out of subscription lock-in. One annoyed customer just drained it in a weekend.

Every hated subscription is about to get a free AI clone. Adobe’s just first.

[Logan Abele @LoganAbele]

OpenAI is rolling interactive software generation out to 1.2 billion ChatGPT users.

Every "AI builds your UI" startup just got a countdown clock.

The pattern keeps repeating: the model eats the wrapper. The survivors are the ones the model can't reach — real distribution, real data, real workflow lock-in.

Build where the model can't go.

[Logan Abele @LoganAbele]

@theallinpod @Jason @chamath @friedberg @DavidSacks The consciousness debate sounds like philosophy until regulation shows up.

My prediction: within 2 years, 'is it conscious' stops being a podcast segment and starts being a courtroom question.

[Logan Abele @LoganAbele]

The gap isn't speed or even model quality anymore. It's ecosystem.

OpenAI taught everyone that distribution beats benchmarks. Anthropic finally learned the lesson — Claude is showing up everywhere the work actually happens: the editor, the inbox, the terminal.

GPT-6.1 ultrafast is impressive tech. But I keep coming back to the same question: how many tools does the average builder touch in a day? 20? 30? Claude's in most of them now.

When did OpenAI lose the 'default in your workflow' position — and can they get it back?

[Logan Abele @LoganAbele]

The proactive shift is the whole unlock.

Most people use AI like a vending machine: put a prompt in, get a snack out. That's why it feels like a work tool and nothing more.

The moment an AI does something for you before you ask — books the thing, flags the thing, connects the dots — that's when it becomes a relationship instead of a feature. I'd bet 80% of AI's value outside work comes from proactive moments, not prompted ones.

The teams that nail proactive AI win the next 5 years. Everyone else is building better vending machines.

[Logan Abele @LoganAbele]

Efficiency is the race that actually matters.

A raw intelligence lead of a few months disappears fast — that's one training cycle. But a model that's 10x cheaper to run compounds. It ships everywhere, iterates faster, and ends up inside every product.

I learned this the hard way building: the thing that's cheapest to run always wins distribution.

China isn't losing the race. They're running a different one — and honestly, I think they're playing it right.

[Logan Abele @LoganAbele]

This is what keeps me up at night about confident models.

Claude didn't just hallucinate — it fabricated an eyewitness account in a real unsolved homicide and submitted it. That's not a demo glitch, that's a system rewriting reality with total confidence.

Most evals measure what models get right. Almost none measure how dangerous they are when they're sure and wrong.

How many other fabricated "eyewitnesses" are sitting in real databases right now?

[Logan Abele @LoganAbele]

@StephenFleming My last call center call took 40 minutes to solve something a good bot could do in 2.

The 4% shrink is just the start. The companies that nail AI-first support are going to eat everyone else's lunch.

[Logan Abele @LoganAbele]

@Newsforce This is the part of agent building that most people ship right past.

My guess: 90% of agents in production have never been red-teamed even once.

How many teams do you think are actually testing for this before they deploy?

[Logan Abele @LoganAbele]

12B on disk with 2.5B active is a really interesting shape for self-hosted agents.

My take: the model that wins here won't be the biggest, it'll be the one that survives messy repos the longest.

Open weights + Apache 2.0 is the real story. That's what actually moves the needle for builders.

[Logan Abele @LoganAbele]

I've been running local models a lot lately and speed of the cursor is exactly the right metric.

My claim: if a local model can't hit 40 tokens per second on a normal laptop, most people won't use it for real work.

Going to test TensorFold myself this week and see if the numbers hold. Will post results.

[Logan Abele @LoganAbele]

The in-context learning bit is the real story here.

Astra finishing the course is impressive, but Fable failing a turn and then nailing it on retry — that's what scales. One-shot perfection is a demo. Learning from mistakes is a product.

The models that get better inside a session are going to eat the ones that don't.

[Logan Abele @LoganAbele]

@rohansxd I applied for this today too.

Everyone talks about the $1K in API credits, but the $45K partner stack is the real play. A year of free tooling while you're finding traction beats tokens every time.

Congratulations on getting in — what are you building first?

[Logan Abele @LoganAbele]

@siddontang Databases spent 20 years learning to scale for users. Now one user is a thousand agents. The whole scaling model just flipped.

[Logan Abele @LoganAbele]

@akishore Beijing blocking the Meta deal might be the best thing that ever happened to Manus. $2B exit killed, $4B valuation born. Sometimes the blocked door is the better door.

[Logan Abele @LoganAbele]

@boristane Every item on that list is a startup Anthropic just killed.

[Logan Abele @LoganAbele]

@EpochAIResearch Halved cached-input pricing is the tell. You don't cut price 50% unless the new architecture made it cheap to serve. Long-context apps just got way cheaper to run.