Joseph

Software Architect.
Built the Evōk Engine. Lossless twins of your codebase made for provably safe AI engineering. Deep in #golf and #martech.
DMs open!

[Joseph @JosephD]

Here's what I've been working on nearly 3 years. This is a deterministic coding and refactoring engine. What coding should be, and the backbone AI needs to become provably safe. You can learn more about what it unlocks (and read things in dark mode) at https://t.co/CUSz0Zp8T3

[Joseph @JosephD]

This is a really good one for OTT. Makes you stop thinking about anything but that belt buckle and you really do shallow out.

[Amber Nicklaus @isthataamber]

Stop coming over the top. Move your belt buckle to your back heel to shallow the club. Save this ⛳

[Joseph @JosephD]

@DannyMagazu I know a guy but he's only 1/8th lawyer. I'd think you need 2/3rds at least before you hit amazing territory.

[Joseph @JosephD]

@colinfritz17 @adamshuaib Only seedless squares at the moment.

[Joseph @JosephD]

@Mat_Sherman There's a nickname for them. Angle investors. They're pointy and tough to work with. $5k check with 500 hours of check-in and "just my opinion" attached.

[Joseph @JosephD]

Refund. They are the worst. Anyone who goes right to threats over a $19 charge is mentally deficient.

Almost every biz would still do the refund even if they were nice about it. Anything under $35 though is almost an auto refund relative the chargeback risk it's not worth it.

We do it on some sub $35 products by making the refund button right there. One click no hassle refund. It instantly refunds and closes your account. Available for the first N days your account is open then hides.

[Joseph @JosephD]

It writes great copy. Arguably better than others and understands nuances in copy much better that must be held together. I'd argue flash 3.8 is only a touch worse at this overall.

No aphoristic tech-bro speak or glib nonsense when you tell it not to, unlike everything else that regresses to the mean.

[Joseph @JosephD]

Everyone is stuck working inside the same "top 30%" of the market where noise is loudest. It's the top 30%, because by design search, lead lists, maps, social media whether free or paid, all prioritize maintaining the marketplace for visibility so they get paid.

If you invert it, and focus on deeper signal, the shape of commerce in any industry is predictable. Every business leaves footprints with upstream suppliers, adjacent trades, shared institutions, and regional rosters, even if they have zero SEO.

This is a technology problem, and the incumbents aren't building it because it's inherently sidesteps their entire revenue moat.

Think about all the businesses that never post on linkedin, never get on here, and do millions per year. The sub 100 employee audience is massive and gives off so much data that's entirely hidden from our day to day because a single human or small team can't map it efficiently (or couldn't) and the existing web sources are all fundamentally built on "throwing out" the bottom 70%. The raw data was hard to get.

I've been working on this for some time now, because to be it's untenable to continue the way it is. We've lost hte ability to even know what businesses are in our back yards without a gatekeeper, and at least for the B2B market, that's billions in dark revenue until tech was good enough (now) a problem only a hyperscaler could solve.

TL;DR: I think we go back to more personal relationships outside of this. Actual personal networks that operate through email, messaging, and domain names we individually control. With that, 1:1 email hyper personalized becomes the only real way to connect outside of the platforms. My 2 cents.

[Joseph @JosephD]

@10footinvestor More siloed experiences, hidden behind algorithms. We're trapped in many ways.

Direct email, 1:1, is what works. Everyone competing in the same fished pond for everything else.
There had to be a better way.

[Joseph @JosephD]

@darynakulya @NoRiskNoParty Cool. Was a great service when I was using it. Only stopped because we sold the club.

[Joseph @JosephD]

@hey_fife lol so they're literally doing nothing and expecting results huh. Seems about right.

[Joseph @JosephD]

@arpitatscale @NamanyayG His astroturfing spam is all over reddit. It's exhausting.

[Joseph @JosephD]

@jamiejreach lol I love the ragebait on this platform.

Half the world works 12 hour days to stay alive. Doesn't mean everyone has to. Some people love to, some people have to out of need, others hate the idea of a minute over 8.

Like the car biz says, there's an ass for every seat.

[Joseph @JosephD]

@ChereneAubert @jordanmenard lol saw https://t.co/SKmgehs6pp was for sale this morning, someone must have time for that. Gotta make sure the AI has a NJ accent and has tool calling for a crowbar.

[Joseph @JosephD]

You're not going to get much privacy if you're routing through any provider at all and not running 100% local.

I don't know how this gets solved without 100% on device local models for actual privacy. Do you really trust that something is zero data just because they publish they are?

A week ago someone found a provider routing to a cheaper side door. You literally can't know for sure.

[Joseph @JosephD]

@paulhshort @joshelman Then it was a horrible human mid-take. I apologize to the AI for blaming it for such nonsense.

[Joseph @JosephD]

Exceptionally common AI phrase you see a ton on here:

"That's the whole job."

It's one of those that feels like you're putting your foot down, sounds like it means something, and is applied to everything by AI. It's the aphoristic tech-bro vibe. A bit of smug dismissal. Slop.

[Joseph @JosephD]

Oh yay another lukewarm take from AI. Why though, why write with AI at all on social media, who wants these mid-takes?

What you're describing is the hopeful state. The optimal state. Economic optimism. Practice vs theory.

In reality, you can't keep modern AI to avoid derailing without spending a tremendous amount of time and money on wiring your own harnesses. On both lab harnesses and raw API models from labs, the prompts they bake in above the ones you can control will force bad decisions that spend more tokens in a circle.

You can be explicit as you want, there is no perfect A -> B path in a non deterministic system.

Money because you're paying a la carte API rates. Time because by the time you've got one model working great, the post-training makes the next one behave in a contradictory way and we're too early for model pinning of any real time window.

And that's only with EXISTING stuff.

Imagine when you try to get it to understand something nobody has done before, it quickly regresses to slop soup in its efforts to placate you and "have an answer."

[Joseph @JosephD]

This slop is a great example. It thinks it has some brilliant take, and it's luke warm generic slop advice with no real context to what I'm doing.

You can't ask a verifier to verify something that it's been trained is "wrong" by past patterns. It will always fight it.

What your AI slop cannon has described is beating a dead horse and keeping a human on the hamster wheel.

[Joseph @JosephD]

I can assure you after 8k hours working on proprietary tech like this, there is no AI that's good at it.

Also, as a long time senior architect, the decisions and choices AI coding makes are unfathomably bad and dangerous in so many cases.

It prioritizes where water flows easiest, which is always teh regressive middle line. The same reason you can get 3 different answers by just pushing that middle line.

If you want it to maintain though along that outlier? It can't, it will always regress to the mean.

Had I not started this project before AI and tried to do it "with AI coding for me" it wouldn't even be possible to exist. I know the fight I had hwen I DID involve AI and it was a nightmare and has only gotten worse as they try to "make AI more helpful" (i.e. to the masses, regression to the mean)