JoshXT

6.6K posts

JoshXT banner
JoshXT

JoshXT

@JoshXT

Founder, CTO & Software Engineer @ https://t.co/psDIioAvYS, https://t.co/q7ENklCk4W, https://t.co/AAgBOldMaL, https://t.co/142r1MAKUa . Forever curious & building open source.

United States Katılım Kasım 2010
1K Takip Edilen6.8K Takipçiler
Sabitlenmiş Tweet
JoshXT
JoshXT@JoshXT·
In early February 2022, I was down with severe case of food poisoning. The most major turning point in my life should have probably been having kids, but instead, it was almost being claimed by food poisoning. Some people lay in bed praying or accept what they believe to be their fate, I laid in bed promising myself that I would do better and be better if I made it out. I have no intention of ever breaking that promise to myself. We're almost at 2 years since then, I have worked 80-100+ hour weeks every week since then to improve my knowledge and abilities to enable myself to build a better future for me and my family. It is paying off in ways I could not have even imagined 2 years ago and I am so far from done. I'm the best me I've ever been, I am doing everything I can to be able to say that every single day.
English
53
9
217
119.6K
JoshXT
JoshXT@JoshXT·
@paulg I've been sitting here for the past 10 minutes wondering what social errors are and how many I must commit when I interact with people. Ignorance is bliss though.
English
0
0
0
65
Paul Graham
Paul Graham@paulg·
Even though nerds are socially awkward, its actually easier to hang out with them than with smooth people, because standards are lower. You don't worry that you might be making social errors; all of you always are; so it stops mattering.
English
174
134
2.5K
85.2K
JoshXT
JoshXT@JoshXT·
@mcuban It depends on what you want out of it. If it answered the same way every time for coding problems, it wouldn't be useful to me, it would be too rigid and not creative enough to solve new problems.
English
0
0
0
32
Mark Cuban
Mark Cuban@mcuban·
I’m coming to the conclusion that the biggest challenge for Enterprise AI, and AI in general , as of now, is that it’s still impossible to make sure that everyone gets the same answer to the same question, every time. Which is a great response to the doomers. AI doesn’t know the consequences of its output. Judgement and the ability to challenge AI output is becoming increasingly necessary, and valuable. Which makes domain knowledge more valuable by the second. Am I wrong ?
English
1.2K
205
3K
426.6K
JoshXT
JoshXT@JoshXT·
@creepydotorg The real Darth Vader story is still playing out.
English
2
0
427
25.7K
JoshXT
JoshXT@JoshXT·
@sudoingX Q4_0 on KV has been perfectly fine for me with Qwen3.6-35B-A3B with unsloths Q3_K_XL quant. I'm running that on a 5090, a 4090, and a 3090 in different machines. It is crazy how good it is.
English
0
0
1
303
Sudo su
Sudo su@sudoingX·
nobody wants to hear this but on a single rtx 3090 or smaller consumer card, q4_0 kv cache is the right choice. that is the default i ship in every benchmark and every live session, and i have not seen a real workflow degradation yet across thousands of inference calls i have personally run. q8 kv cache is technically better quality and holds attention precision better at long context, but on 24gb it does not fit at 256k context without slowing the model 3x. q4 does fit, comfortably, at the speed the model was designed to run. most of my audience runs on a single 3090 or multi 3090 setups. some of you are on 8gb cards, where qwen 9b at q4 with q4 kv cache is the right entry, fits and chats fluently. there is no shame in the consumer-tier stack, the shame is in not knowing what tradeoffs you are actually making. people saying q4 kv cache is dogshit are usually running on hardware where the trade off does not bite them. on consumer hardware it is the right choice for almost every workflow. test it on the work you actually do, that is the only benchmark that counts.
Sudo su@sudoingX

if you are starting local ai from zero, i'll drop the full guide in the open source release but the short version: any modern desktop with a single rtx 3090 or 4090 (24gb vram class), llama.cpp built with cuda, qwen 3.6 27b dense at q4_k_m gguf from unsloth, and hermes agent or opencode as the harness driving it. ngl 99 to offload everything to gpu, c 262144 for 256k context, fa on for flash attention, and the q4_0 kv cache to fit it all in 21gb. the rest is the prompt. that is the whole stack.

English
17
8
185
16.6K
JoshXT
JoshXT@JoshXT·
@iliketeslas I'm clicking critical every time until they stop blocking my nav when I disengage.
English
0
0
1
43
i like teslas
i like teslas@iliketeslas·
Highly suggest Tesla just assume the reason we disengage is navigation. There is no other reason why to disengage
English
141
3
471
22K
JoshXT
JoshXT@JoshXT·
@Austen You just directly connect to the third party API?
English
0
0
0
52
JoshXT
JoshXT@JoshXT·
@catalinmpit It depends on what you're doing. Claude is significantly better at UI still, but got-5.5 in Codex is good at most problems I've given it besides that.
English
0
0
1
97
JoshXT
JoshXT@JoshXT·
@thsottiaux This should just be default behavior. I never want to ask AI to do something and it be half assed before I get a response.
English
0
0
0
59
Tibo
Tibo@thsottiaux·
You can now keep codex going for days. With GPT-5.5 it will build an entire OS kernel for you if you ask, or find critical bugs in a codebase, or optimize your database schemas, or… the options are endless.
Felipe Coury 🦀@fcoury

/goal also lands in Codex CLI 0.128.0. Our take on the Ralph loop: keep a goal alive across turns. Don't stop until it's achieved. Built by my co-worker and OpenAI mentor Eric Traut, aka the Pyright guy. One of the GOATs I get to work with daily.

English
337
255
5.4K
688K
JoshXT
JoshXT@JoshXT·
@jasperdevs This should just be default behavior. I never want to ask AI to do something and it be half assed before I get a response.
English
0
0
3
282
jasper
jasper@jasperdevs·
CODEX FINALLY ADDED BUILT IN RALPH LOOPS it apparently can run for multiple days straight if you let it, until the task is finished its called /goal and to enable it you have to go to config.toml, ctrl f to find [features] then add goals = true
jasper tweet media
Tibo@thsottiaux

You can now keep codex going for days. With GPT-5.5 it will build an entire OS kernel for you if you ask, or find critical bugs in a codebase, or optimize your database schemas, or… the options are endless.

English
36
58
1.3K
138.9K
Shaw (spirit/acc)
Shaw (spirit/acc)@shawmakesmagic·
Hey @nikitabier i'm not kidding that 90% of the comments to my posts are AI reply guy agents that didn't ask my permission If you wanna find them literally just go into any of my comments and tell me who is real and human Please I'm begging you
English
71
6
406
112.7K
JoshXT
JoshXT@JoshXT·
My kid is using AGiXT to interactively make a game with his AGiXT agent in an entertainment app I made for my kids.
JoshXT tweet media
English
0
1
3
134
JoshXT
JoshXT@JoshXT·
@jeremyjudkins_ You forgot to mention the complete lack of acceleration. Last time I got in a gas car, I put my foot to the floor because it wouldn't go and the dumb thing spun tires instead of catching traction and just taking off immediately. 0-60 in 30 seconds vs 1.99 is tough.
English
0
0
6
207
Jeremy Judkins
Jeremy Judkins@jeremyjudkins_·
POV: You are a Tesla owner attempting to drive a gas powered car.
English
39
16
323
25.6K
JoshXT
JoshXT@JoshXT·
@wholemars It wouldn't be so bad if it didn't completely cover navigation.
English
0
1
2
45
Whole Mars Catalog
Whole Mars Catalog@wholemars·
I drive you assholes everywhere you want to go all day and you’re too lazy to tap once on the screen to tell me what I did wrong? Really??
Whole Mars Catalog tweet media
English
91
28
740
26.5K
Buddy
Buddy@StanleyDav8740·
@llmdevguy @Protorikis Absolutely false. Qwen3.6 35b is hot garbage by turn 3 in any agentic harness
English
1
0
1
392
Mateusz Mirkowski
Mateusz Mirkowski@llmdevguy·
🙌Qwen 3.6 35B vs Sonnet 4.6 - which one is better? @Protorikis made an interesting comparison between these models (video in comments). TLDR: - Similar time to complete the task - Similar code length - Qwen 3.6 delivered a slightly better solution So local model match Sonnet 4.6. Yes, you can have Sonnet 4.6 at home. 😇
Mateusz Mirkowski tweet media
English
14
2
82
9.9K
Brett Adcock
Brett Adcock@adcock_brett·
In the last 120 days, Figure scaled manufacturing 24x - from 1 robot/day to 1 robot/hour We will manufacture 55 humanoid robots this week
English
335
408
4.3K
442.3K
JoshXT
JoshXT@JoshXT·
@mgirard @coooooooopppppp @Tesla @grok I can at least set my max speed in my HW3 S. Cybertruck just gets me speeding tickets if I don't disengage or go from standard to sloth to get the nice hard brake feel.
English
0
0
1
53
Tesla
Tesla@Tesla·
Following future rollout of FSD V14 Lite for HW3 vehicles in the US, we plan on expanding V14 Lite to additional international markets. This update ensures that HW3 vehicle owners will continue to benefit from ongoing software updates. Since international rollout is subject to several factors (completion of technical verification, regional adaptation & relevant regulatory approvals), we can't provide definitive dates at the moment, but will provide updates on a rolling basis
English
1K
2.3K
12.4K
1.9M

Keşfet

@thedooberhead @paulg @mcuban @creepydotorg @sudoingX @iliketeslas @Austen @catalinmpit