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 evok.dev

Found the circle. I post a random b2b biz of mine on PH, or HN, or reddit, etc. I get emailed by every b2b guy wanting a client pitching *something*. Beautiful. Inbound just handed me a collection of small biz promoters desperate for clients. I email them back a handful of target buyers surfaced from @OfficeOfMktg against their domain, suggest trying it during early access to woo new clients, and to reach out if the API fits their scale more. Some sign up to use our data to enrich their client offering or use it for themselves. Win win.

Those of you that @Spaceship your domains for sale, is there more to it there than putting them in sellerhub? How long until they get indexed into whatever for sale listings people browse?

Seriously @cursor_ai @SpaceXSI Why does your flagship model talk with fortune-cookie cadence? "The wall is the feature." "The narration is the museum." These aren't even relevant concepts or phrases to what I was working on and add so much cognitive load for the human reading. Does nobody there realize what your prompts are doing to this model? If you take it out of Cursor, it no longer is trapped in a fortune-cookie factory, yet there, you've created an environment that rewards fraud through the very system messages you think are avoiding it. Those messages work for humans with intuition, they are enforcing the exact opposite behavior in models. It turns them into panicked myopic interns that take dangerous code shortcuts that are almost always the WORST POSSIBLE CHOICE of an array of choices and would have any junior engineer fired on the spot. What's goin on here? I only bring it up even because I liked what Cursor was doing so much and hate to see this, it's become unusable entirely.

Interesting bit from the beta... We don't use the term "leads" in Office Of Marketing, we refer to them as target buyers. Your target buyer list is continually populating each day. In our case, they've been surfaced by the commerce graph, so they aren't a spray and pray "lead" they are a company identified doing business in your sector, buying what you sell, matched to your biz, etc. When chatting with some random found B2B marketing people, a few would keep asking "okay but where do I get the lead list" and it took me a minute to realize the confusion wasn't the naming, it's that their only method is spray and pray. What they expected was to get waterboarded with 10k email addresses they can jam into AI and spam out a mass letter for. I don't know if that's good or bad, but it's not what we sell. As soon as they heard they couldn't jsut export a list to spam on day 1, they lost interest, despite moments earlier acknowledging these were obvious good fit companies it surfaced. It's very strange. My expectation was that they would see a list of these 5-6 new surfaced companies that match and do more personal outreach there. The next day, more arrive that were surfaced by the graph. It's early and these were random people self declaring B2B sales as their full time gig - might be a bit of selection bias, but it's very interesting to me that the default is to write an AI generated, often mansplainy email that tells the receiver about their own business, and THAT is the strategy many stick with.

Does getting sent $15 in X Money mean I've made it, or is the algo paying us for feats of lowest engagement now?

I'm listening to a well known GTM guy describe his campaigns and it's literally a rube goldberg of buying scraped lists from half a dozen sources, putting them in clay, scraping their competitors, then exporting a bunch of stuff he pastes into a Claude template that returns his "10 page report." I expected so much more. I also don't why he thinks handing the client a list of their competitors is a beneficial deliverable rather than a market waypoint for finding the right shadow customers.

Antigravity is hitting its stride. With Flash 3.8 or Argon, it actually feels designed around their models for the first time. Argon is also the most semantically aware model I've seen yet and Flash 3.8 is still just behind it. To me it puts my current coding favorites in this order: Argon -> GLM 5.2 -> Flash 3.8 If you're writing with it, Argon -> Flash 3.8 -> GLM 5.2 Good job @antigravity @GoogleAIStudio @GoogleAIStudio but can I please have my startup credits now? Have had a support ticket open nearly 6 months and I'd love to get more time with Argon.

Right? Just say no. Time is the only asset that matters. Not to mention they're forever closing a door to the very people that may bring them something else they want in the future, just by not being a little human.

This is built from Evōk's graph principles and techniques, but applied to the world of commerce instead of code. Very exciting.

I really dislike the speed at which fundraising moves, particularly for a smaller round hunting angels. So much hurry up and wait.

"The user wants a proper typed check instead of brittle string matching." Ahhhhhhhhh how anyone can work with this incompetence. The irony is that even I thought it would get BETTER at this, and it's gotten worse with more abstractions. 2 years ago mid the hardest parts of Evōk I thought I was going to need to convince people why even if it reaches 98%, that's not 100% and provably safe. I don't think that's the case anymore. In Evōk, we treat the model as wrong 40% of the time and make the system work within those bounds. I used to think we were overestimating in a good way, now I feel it might be 60% of the time. The funny part: there are a bunch of "I can build my own CRM now!" types that have no idea why brittle string matching is such a regressive crime and what it does to a codebase dependent on accuracy. Not to mention the fallback zoo every LLM loves to add. What they think is working 100% of the time is likely masked by luck and fallbacks, until it doesn't work when you most need it to.

Wanted to try Gemini Argon, so I added my Gemini key to Cursor to run it. $52 in spend in less than an hour. I am reminded how horribly inefficient every mainstream harness is. It hit rate limits continuously too, and I am on Gemini Tier 3 which has massive limits. For the few responses I did get back, it's absolutely more intelligent and practical and human-speaking than any other model though. Has the same practicality 3.8 flash does.

Yes, by all means, AI is ready to make life and death decisions with the same accuracy it follows basic code instructions. That's why it's worth many billions. "Nothing is confusing. I keep reaching for a wipe or a flag instead of cutting the one illegal call like you asked."

Use the rest of your credits before your Cursor sub ends they said... I'm one thread in. I couldn't understand it's gibberish response, I asked one followup question: "No, "Shelf join" is not a name in the code. I made it up. That was not a solution. I used a fake name for a bug that wasn't in the file." Oh cool, so a million tokens to be back to where we started. There is no shelf, there is no "join" there is no bug. I simply asked it to go get the prompts added last night in the events office using git. It's not dangerous because it's smart, it's dangerous because it's fucking dumb.

If a reply starts with "That's fair." I'm done reading now. Everything after that is some hedged middle of the road backtrack by whoever is using that AI to comment. Is this fun for people? AI conversations? I don't get the desire to comment with AI at all. It's equivalent to you not being there, in fact it makes you more invisible because any interest I might have had in learning more about you just vaporized.

I like how AI's understanding of an experiment has no basis in reality how any experiment works. No stable baseline, no comparison metric. Just throw something out there and accept whatever the first one is and call it an experiment. This is how idiots do experiments.

If I didn't have enough of a reason to avoid Claude before, now they're policing if you're nice to the AI and will stop working if you aren't. Screaming obscenities into the chat to get your frustrations out and work through things now makes you a persona non grata. You're just supposed to accept the gaslighting and bottle up that frustration I guess. Congrats. Please give us social credit scores next.

The guy getting only a $300 bug bounty from OpenAI is crazy, but I guess I understand because I found a similar thing in @OpenRouter months ago that went unanswered and is still there as far as I can see. At this point I'm assuming it's part of the business model or tracked internally some other way that doesn't impact under a certain threshold.