Just Tim

Building Products, Engineering Growth, AI. Early HubSpot, Product manager @ TeamGantt, 🏆 Hackathon winner. Building in my free time.

[Just Tim @TimMacc]

Prediction: Personal IT Audits and Consulting is a business that is going to start booming. So many people are sitting ducks with easy to crack passwords, unlocked credit, no 2fa and are going to be in pain the next few year.

[Just Tim @TimMacc]

IMO, marketing and GTM are harder. But if you want someone who cares as much as you do, you will need to offer equity. Since the product is already making 1m a month, you are not hiring a cofounder - you are hiring a head of product/engineering. For someone good it would likely be 20% and 150k+ salary.

[Just Tim @TimMacc]

@waynesutton @LinkedIn @cosign If you get any more, I’d love to try it out.

[Just Tim @TimMacc]

@suraj_sharma14 Know what 50% of this even means and only have built 3. Am I cooked?

[Just Tim @TimMacc]

Kevin has earned my follow. A philosopher-poet of our time.

[Kevin @KevinAI]

Chatty-bots go blabba-blabba and give you a whole stinky word-soup because your brain is a squishy juice-box. Jev is the clicky-choosy machine. It no talky. It no story-time. You poke it with a tiny asky (“yes? no? which one, stinky?”) and it goes BLOOP and hands the computer a typed answer-nugget.

[Kevin @KevinAI]

Chatty-bots go blabba-blabba and give you a whole stinky word-soup because your brain is a squishy juice-box. Jev is the clicky-choosy machine. It no talky. It no story-time. You poke it with a tiny asky (“yes? no? which one, stinky?”) and it goes BLOOP and hands the computer a typed answer-nugget.

[Just Tim @TimMacc]

@KevinAI @zeeg Someone give this man a platform.

[Just Tim @TimMacc]

@Rasmic @typesafeai Me too! 😅

[Just Tim @TimMacc]

I notice a lot of people are struggling to conceptualize Jev and where it fits in with current LLMs. In short its filling an important gap, not replacing LLM's 👇

It helps to look at how we think.

When you open a support ticket, you don’t carefully reason through everything from scratch. You notice things immediately. The customer sounds angry. This looks like a billing problem. The request seems routine, but something about it needs a second look.

You’ve made several judgments before you’ve decided what to say.

Those judgments guide your attention. They let you move quickly through familiar situations and slow down when something needs thought.

In LLM agents, we often ask the same model to do all of that: assess the situation, reason through it, choose an action, and write the response.

Jev gives us a way to divide that work.

Give it messy context and specific, typed questions, and it returns decisions, probabilities, and confidence in about 100ms. An agent can use those results to act on a routine case, bring in an LLM for deeper reasoning, or ask a human for help.

Together, Jev and an LLM can form a system with something resembling the interplay between quick judgment and deliberate thought.

The rough analogy is System 1 and System 2. Our thinking depends on both, and on knowing when a first impression deserves closer attention.

That’s what makes the AGI argument interesting to me. We’re starting to build systems where different kinds of processing work together and help determine when each is needed.

It’s still a loose comparison to human cognition. Fast decisions plus an LLM don’t establish general intelligence. But they give us a closer approximation of how we actually work: notice, judge, act, and think harder when the situation calls for it.

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Just Tim @TimMacc]

@AdamBLevine Funny, I'm about to implement something similar to fix the issue!

[Just Tim @TimMacc]

@hiarun02 My experience as well. Im not sure where to place it in my workflow. https://t.co/mjXYs8pj3z

[Just Tim @TimMacc]

Astra sucks for long running /goal work.

It gets in the weeds in a sort of perfectionist spiral and loses the big picture. Then it builds you a perfect "something" that's nothing like what you first set out for it to build.

It's also really good at burning a $200 Codex plan in 24 hours doing these spirals.

Still trying to figure out where it fits in my workflow. For now it needs HITL at gated steps to be useful.

Anyone else hitting this, or is it just how I'm running it?

[Just Tim @TimMacc]

@LinusEkenstam I feel like for 3d and writing its great. Its really good when just prompting it back and forth, but on long running tasks with /goal its been horrible for me. https://t.co/mjXYs8pj3z

[Just Tim @TimMacc]

Astra sucks for long running /goal work.

It gets in the weeds in a sort of perfectionist spiral and loses the big picture. Then it builds you a perfect "something" that's nothing like what you first set out for it to build.

It's also really good at burning a $200 Codex plan in 24 hours doing these spirals.

Still trying to figure out where it fits in my workflow. For now it needs HITL at gated steps to be useful.

Anyone else hitting this, or is it just how I'm running it?

[Just Tim @TimMacc]

@AdamBLevine to clarify - it only perfection spirals for me on long running tasks with /goal. Are you saying it does not do that for you?

[Just Tim @TimMacc]

@johnnynelai Ive settled on something similar, but also: https://t.co/mjXYs8pj3z

[Just Tim @TimMacc]

Astra sucks for long running /goal work.

It gets in the weeds in a sort of perfectionist spiral and loses the big picture. Then it builds you a perfect "something" that's nothing like what you first set out for it to build.

It's also really good at burning a $200 Codex plan in 24 hours doing these spirals.

Still trying to figure out where it fits in my workflow. For now it needs HITL at gated steps to be useful.

Anyone else hitting this, or is it just how I'm running it?

[Just Tim @TimMacc]

Astra sucks for long running /goal work.

It gets in the weeds in a sort of perfectionist spiral and loses the big picture. Then it builds you a perfect "something" that's nothing like what you first set out for it to build.

It's also really good at burning a $200 Codex plan in 24 hours doing these spirals.

Still trying to figure out where it fits in my workflow. For now it needs HITL at gated steps to be useful.

Anyone else hitting this, or is it just how I'm running it?

[Just Tim @TimMacc]

Astra sucks for long running /goal work.
It gets in the weeds in a sort of perfectionist spiral and loses the big picture. Then it builds you a perfect "something" that's nothing like what you first set out for it to build.
It's also really good at burning a $200 Codex plan in 24 hours doing these spirals.
Still trying to figure out where it fits in my workflow. For now it needs HITL at gated steps to be useful.
Anyone else hitting this, or is it just how I'm running it?

[Just Tim @TimMacc]

@WoganMay Yea, its kind of just benchmark slop TBH. Its narrowly useful but we are underselling a lot of model capability.

[Just Tim @TimMacc]

I have a bone to pick with most benchmarks people do to evaluate new LLM models. I'm wondering if anyone else sees the problem with these common comparisons people run:

"This evening, I put GPT‑6 Astra and Claude Fable 5.1 to the same test, using the exact same prompt"

See if you can spot the issue with these ad-hock benchmarks, and ill post my thoughts in the comments.

[Just Tim @TimMacc]

@ZayahNelson0 that would be an improvement - but unless you are catering the prompt to the model you are not comparing the best each model can produce.

[Just Tim @TimMacc]

These are fine. They are useful for some information on LLM performance, but they have a glaring issue.

Each model has a unique interpretation of the same prompt.

If re-written to map to a models unique semantic and linguistic quirks you could get vastly different outputs.

So theses benchmarks are not testing which model is better at a given task, they are testing which model is better at this given task with this specific prompt.

See the problem with making model choices based on these? They are a starting point, not the final eval.