Software Architect. Built the Evōk Engine. Lossless twins of your codebase made for provably safe AI engineering. Deep in #golf and #martech. DMs open!
Palm Springs ♻ Miami
- Tweets 4,023
- Following 3,947
- Followers 1,274
- Likes 6,432
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
5
2
34
3,159
I don't know anyone at @cursor_ai so.... @elonmusk please do not allow this to regress to slop soup to appease the lowest common denominator.
The purchase was clutch, makes so much sense, but this old style prompt jamming has got to be phased out asap.
10
I really want to like Cursor, they've done some nice things, but this bump to 4.7 hurts. The voice side is awful suddenly, it gets every third word and overwrites itself.
Then, the same problem still (like over a year) exists where they don't' bother telling the model that it's voice type, so the model takes everything literally including spelling and however voice type decides to punctuate.
Try saying a company name, and it turns into some incoherently slop that then the AI binds aggressively to and starts a rabbit hole of more slop.
Like how come every small time harness has figured this simple thing out but Cursor hasn't?
Just tell the model to be flexible and look through for intent relative the ongoing chat rather than treating obvious voice transcription issues as literal commands.
Just wrap the incoming user message in a [Voice Transcribed. May have typos. Read for intent relative the recent conversation. ASK for clarity if gaps exist.]
Would save so much damn aggravation.
(And the only reason I can imagine you would NOT do this is if you're encouraging pointless token burn)
1
2
65
lol, and in the Cursor system prompt... baked in to every call: "Bias towards not asking the user for help if you can find the answer yourself." (What it hears: make shit up and do whatever to reach 'done')
And this gem as if nobody @cursor_ai ever tested what this does to its bias:
"Look past the first seemingly relevant result. EXPLORE alternative implementations, edge cases, and varied search terms until you have COMPREHENSIVE coverage of the topic."
This doesn't do what you think it does.
It turns it into "everything is a nail" mentality. It WILL find an alternative way that disagrees with the users commands, which is precisely the behavior it exhibits on every task.
Then I ask the model and this is what I get back:
"It is the exact same myopic, lazy hack: hardcoding a band-aid so the one test in front of my face passes"
Come on guys. Stop depending on the model intelligence and build some architectural intelligence.
11
Grok 4.7 needs an award for "most able to sit on a fence and say nothing."
You give it an intent, ask a question, and you get back the flattest middle of the road NON choice. You have to be a forensic investigator to figure out whether it is speaking positively or negatively about anything it wrote.
Like they stripped all bias to the point where it's a fucking null character that can't even align with your intent because it's so scared to pull the fencepost out of its ass.
1
3
70
Me: Look at the intake API. Every 3rd call fails. This is likely a concurrency issue. Call 3 must use data from call 1 and 2. They must run as 2 concurrent, followed by the 3rd call.
AI: I changed the timeout on the 3rd call from 10 seconds to 90 seconds so it has time to complete.
Me: That is a worthless patch. The API takes 3 seconds max you fool, adding 90 seconds does not solve concurrency. Check how they fire now, this is an issue in how you're processing the upstream command. It has nothing to do with process time, it's ORDER OF OPERATIONS and properly awaiting a result.
AI: You're right! I've added a fallback so the third call can fail gracefully.
Me: (looking around the empty room seeing if anyone else is seeing this shit... unwinding the last changes... )
What ORDER are you firing the 3 calls. If the third call is not waiting for the first 2 to finish, it WILL NOT COMPLETE.
AI: "I've found it! I am calling all 3 calls in parallel every time. Let me make those sequential..."
Proceeds to make ALL CALLS SEQUENTIAL --
Me: You went out of scope. Only the third call, the one LITERALLY MARKED #3, must be held until 1 + 2 return. You NEED INFORMATION FROM 1 + 2 YOU. YOU CAN NOT DELIVER 3 WITHOUT INFORMATION FROM 1 AND 2.
AI: "You're right! Let me..." (Proceeds to write NEW code, not edit what it did, but a NEW function for this, right next to the old one) ... "and I've made the original function now return null."
Me: Let me get this straight, so your solution is to deprecate the actual function, replace it with a new one you finally write correctly, and LEAVE the trash old one there as a dead end? Why would you not just fix the intended function? What is the technical reason you are choosing the worst possible path?
AI: "You're right!....
--------
There are so many "works once in a demo" apps out there that are getting funded like crazy that don't even understand why this conversation is infuriating beyond the back and forth. It's the regressive thinking, the "patch at any cost" mentality, the "doesn't matter, more tokens will solve it later" thinking.
Imagine the number of people that accept the first answer.
1
1
87
You guys are also all screaming endless obscenities into AI right? Not just me?
Tried the newest models, I swear this is the biggest scam on earth. They are building like AI builds, abstraction after abstraction to the point where the thing can't do a single task without breaking a hard rule because it's stacked on stacked shit just there to eat tokens until it gives you an answer that's 40% wrong even when it's right.
With all the money and talent they have, this is the upper end?
We're fucked.
7
126
It's satisfying to see these guys trending on twitter today getting called out for this very bullshit.
Glad someone else finally looked at the code.
There is a coding harness out there that has millions in funding, open source, and I don't think anyone in their life has looked at the code.
It's the biggest rube goldberg piece of shit I've seen in a long time, backed by (i.e. funded, obv not on the ground) genuinely intelligent people. I don't get how that happens.
If you think AI overhypes, you should see some of what they market vs what exists in the code. It's absolute trash built on brittle string matching and yelling at the model. They wrap that up in terms that aren't even accurate to what it is, and in reality, it's 3 incoherent bulletpoints in an MD file.
The sad part is, they'll do fine, because it's "good enough" despite being hyperbole filled nonesense, but will those investors ever make their money back? Did those investors know they were buying this level of trash? Do they care?
1
1
211
By the end of today, everyone will be talking about how they've integrated Jev and are doing 100x the quantaflops with their zigbee reduction methods over latent space and how now "decision making is dead."
Meanwhile, in a grocery store this morning, a lady on her phone said her "I put all my recipes on the computer with Chat GBT yesterday it was fun!"
1
2
149
Same writing / site copy task. Tried across a ton of models. The only ones I'll be using for any writing tasks for the near future:
- Gemini 3.8 Flash
- Gemma 4
- GLM 5.2
- GLM 5.3 Flash (5.3 regular is noticeably more mid)
Notable runner ups:
- Kimi K3
- GLM 4.7 Flash
- Qwen Max
2
4
191
What models do you think had the best writing voice?
(Or, the voice you had to fight the least from being "words for the sake of words" nonsense?)
I feel like we need to go back and freeze those based on their training data sets and make them lightweight writing models.
It's just never been as bad as the writing is now. It's all this comparison cadence, via negativa in all forms, constant heros out of banal statements ("The row is the reason"), just all slop.
Even with explicit instructions / skills to bring it towards a different voice, it treats it as an accounting task rather than a creative one, and just loses all intent in the process. Just word salad on these newer models.
1
1
3
105
I tried a new "writing" app I saw got funding. Wanted to see what it did as I've had it bookmarked for a month now. So I gave it a bunch of data on our tee sheet.
This is the gibberish it spit out:
"The desk, the public book, and every partner write the same record. Dining, carts, and booking attach through the club key issues. A booking write is a real reservation. A cart pipe and a dining pipe stay on their own surfaces."
Who is this written for!?!?
I can not understand the meaning behind any of that, and I built the software.
This is a $49 - $149/month product they are selling and raised 1m. I can only imagine their target customer is a goat.
2
1
3
159
I love when a tool ends up being invisible and so useful at the same time. I spun up cursor for a task (hadn't used it in a while so nothing was customized), just attached my one mcp locker link, instantly knew what I wanted done and how to do it.
I definitely think MCP become the "deliver an outcome" method from any given app. Apps get narrower into verticals exposing a delivered outcome via their MCP. Then it's YOUR agent, using whatever tools it needs not itself, but delivered as done.
2
1
2
125
This thread is a great example of everything wrong with golf.
Almost all of them put fault on the golfers or external forces.
Everyone assumes status quo (the industry as it does) does it that way for a reason... They are literally making the case that golf is perfect and it's the players at fault.
You see this SAME level of bro-science in the pro shop, brought on years of doing what the last guy did and not having enough people around them to call their bullshit.
If you use basic math and experiment standards (like maintaining a constant - didn't we all learn this in school?), you can eliminate variables that actually cause the problem with slow play. I'm willing to bet less than 10% of course operators nationally have ever thought about it once. Why would they, it's only their main revenue stream.
Instead, blame the players. Blame carts. Blame music. Blame politics. You name it. It CAN'T be the course ever.
So, they just continue screaming into the void. Courses lock down things creating a rigid relationship, taking away amenities they think are the problem, creating rules and more rules...then when business slows from that, well it's the golfers fault once again and the cycle continues.
Most courses are discounting and trope-believing themselves out of business.
Texting with a friend this morning about the course he works at (GM) and we were talking about pace.
This was a course played a lot growing up.
It was always a 4 hour pace of play for decades. Now its 4:35
10 Minute tee time intervals are the same as they always were.
The course has not been lengthened.
The course generally speaking, hasn't changed in 40 years.
Why is it now 4:35?
1
1
115
Had an appointment yesterday where I checked in on paper, then the person had to handwrite on paper, another person manually updated the computer from that paper. The paper was then shredded.
Meanwhile Twitter would have you believe I floated in on my hoverboard, they scanned my iris, instinct booked 3 flights for me real quick because I thought about a beach, and my agent waited in line and paid while I went to the streudel house for a casual lunch.
1
1
3
110