michael_jay

858 posts

michael_jay banner
michael_jay

michael_jay

@rKaidd

Australia Katılım Nisan 2009
893 Takip Edilen101 Takipçiler
michael_jay
michael_jay@rKaidd·
@sudoingX honestly a.thing you do in this space rn is fkng worth it. i work adjacent to it and still can't work out why more/a the smartest people in my orbit aren't diving in (i feel im very persuasive) the venn of SWE + IaC + AI/ML + actual hobbyist drive is way rarer than it should be.
English
0
0
0
8
Sudo su
Sudo su@sudoingX·
would you watch if i started making short videos of me building?
English
12
0
23
1.8K
michael_jay
michael_jay@rKaidd·
@sudoingX ur qwen3.6_35B recipe, changed my life 4-6wk ago. Sub'ing now.
English
1
0
2
179
Sudo su
Sudo su@sudoingX·
day 6 of the month and i am already 50% through my frontier ai monthly limit. i use these tools to build what i am building for all of us and if you are a builder or your company launched an agentic tool and you want a real tester, i am happy to install it, run it on real work, share honest feedback for some credits. the people who supported me before know who they are. appreciate you. worst interaction so far was cursor. their team reached out offering credits for builders. i never had cursor installed before that dm. after the message, i installed it and emailed with my account as requested. almost 3 weeks of silence since. nobody has even seen the message.
English
5
0
37
3.1K
michael_jay
michael_jay@rKaidd·
@ogrizkov @googledevs Apple and agents. They’ve just put two hardware guys in charge and absolutely going for edge devices and Google know it.
English
1
0
1
291
Evgeny Ogryzkov | Indie AI Builder
Evgeny Ogryzkov | Indie AI Builder@ogrizkov·
@googledevs Still running Gemma-4 locally on Mac Mini M4 16Gb for an agent inside Openclaw. Tested it heavily on iPhone 17 Pro too. It was already fast. Where exactly are they hurrying to?
Evgeny Ogryzkov | Indie AI Builder tweet media
English
7
1
60
13.3K
Google for Developers
Google for Developers@googledevs·
Gemma 4: Now up to 3x Faster. ⚡ Same quality, way more speed. Our new MTP drafters allow Gemma 4 to predict multiple tokens at once, effectively tripling your output speed without compromising intelligence.
GIF
English
167
628
6.1K
817.9K
michael_jay
michael_jay@rKaidd·
@Teknium Won’t lie; gutted, given the hours iv put into building this lol. Jacked to see your implementation! <3
English
0
0
0
8
Teknium 🪽
Teknium 🪽@Teknium·
I have to go out of town for a funeral thru the weekend but I am leaving everyone with one new cool feature inspired by ralph loops and Codex's upcoming /goal feature. If you use /goal , it will start a loop with a supervisor model determining whether the task completed at the end of an agent loop - if it hasn't it will force it to keep going until it's done! Enjoy and have a great weekend. PR: github.com/NousResearch/h…
Teknium 🪽 tweet media
English
130
116
2.1K
365.1K
michael_jay
michael_jay@rKaidd·
@sudoingX RTX 5090 | google_gemma-4-31B-it-Q4_K_M.gguf | ngl=99 c=131072 np=1 fa=on ctk=q4_0 ctv=q4_0 jinja | prompt=164.29 tok/s | predicted=68.31 tok/s
Indonesia
0
0
0
110
Sudo su
Sudo su@sudoingX·
if you run a desktop 5090, 4090, or 3090 and you want to compare, here are the exact flags i'm using: ./llama-server -m google_gemma-4-31B-it-Q4_K_M.gguf -ngl 99 -c 131072 -np 1 -fa on --cache-type-k q4_0 --cache-type-v q4_0 -- jinja same model, same quant, same context window, same kv cache policy. run it and drop your tok/s below with your gpu, i'll amplify the best data points and add them to the benchmark submissions doc. desktop bandwidth should crush mobile here. curious by how much. and we have not pushed the limits yet, this is just getting started.
English
7
2
18
1.9K
Sudo su
Sudo su@sudoingX·
here is first real numbers from gemma 4 31b dense on the RTX 5090 24gb mobile. short prompt: 17.17 tok/s generation long thinking session (1,272 tokens output, full think block + answer): 15.36 tok/s sustained prompt eval: 95 to 165 tok/s depending on length context window: 128k native, q4_0 kv cache, thinking mode on thinking mode burns compute before the answer lands, that's the cost of reasoning. on non thinking models you'd see 20-25% faster generation with the same weights. for context, qwen 3.5-27b dense on a 3090 is 35 tok/s flat, different model, different architecture. the real apples to apples comes next when i run qwen 3.5-27b dense on this same 5090. then we'll know what blackwell mobile actually does vs ampere desktop on 24gb class.
Sudo su tweet mediaSudo su tweet mediaSudo su tweet mediaSudo su tweet media
Sudo su@sudoingX

the 5090 just woke up. gemma 4 31b dense loaded, 128k context, llama-server on port 8080, hermes agent ready on the other side. this laptop has two gpus, the intel i9 integrated for everyday work and the rtx 5090 mobile 24gb for ai. the 5090 sits idle most hours. right now it's spinning up hard, fans sucking air from every direction, my fingers getting cold from the airflow, the entire machine feels awake. next up: speed sweeps across every context size, then autonomous agentic tasks on hermes agent. then direct comparison against the qwen 3.5-27b dense numbers i ran on a 3090 earlier. then qwen 27b dense on this same 5090 after gemma is done. 24gb vs 24gb, different models, same room. and someone anon gave me this laptop. running verified benchmark data for every builder on a machine the internet bought. this is what 2026 looks like when you build in public.

English
5
2
37
29.7K
Nous Research
Nous Research@NousResearch·
Tool Gateway is now live in Nous Portal. No separate accounts, no API key juggling. All you need is one subscription, and everything works. A paid Nous Portal subscription now includes access to 300+ models and a growing set of third-party tools. Launching with: → Web scraping → Browser automation → Image generation → Cloud terminal backend → Text-to-speech
English
254
242
2.6K
2.4M
michael_jay
michael_jay@rKaidd·
@sudoingX I used your recipe from 2-3wks back for my current qwen build; its very performant!! <3
michael_jay tweet media
English
1
0
0
35
Sudo su
Sudo su@sudoingX·
180 tok/s generation on a 4090 with qwen 3.6. if you're on a 4090 and not running this model yet you're leaving performance on the table. 3B active params at that speed is insane for agentic coding. thanks for the data @ErdalToprak, adding this to the comparison sheet
Erdal@ErdalToprak

- model: Qwen3.6-35B-A3B-UD-IQ4_XS.gguf - GPU: RTX 4090 - CUDA, f16 KV, flash attention on - n_gpu_layers=999, threads=8, batch=256, ubatch=256 - Prompt-only, 512 tokens: about 4995 tok/s - Generation-only, 128 tokens: about 180 tok/s - Mixed, 4096 prompt + 128 gen: about 2700 tok/s effective combined throughput - 512,0: 4976.8 to 4994.8 tok/s - 0,128: 179.36 to 179.95 tok/s - 4096,128: 2700.06 tok/s x.com/ErdalToprak/st…

English
28
17
229
25.8K
michael_jay
michael_jay@rKaidd·
@simcity99 wild to hear my 2014 obsession on x. Fully presumed that actors were DM'ing high value users on rcforums, vetting, inviting to private telegram groups, ToR repos and building out "BF for the battlespace"
English
0
0
1
104
simcity
simcity@simcity99·
rc-xd isn't real because nobody wrote betaflight for ground vehicles ukraine scaled fpv drones by riding open source stm32 firmware and commodity hobby hardware rc cars have the same supply chain > turbo ecu is betaflight for ugvs > turbopilot is openpilot for rc vision ai, not gps waypoint ardupilot slop
English
11
18
239
15.3K
Artificially Inclined™
Artificially Inclined™@Art_If_Ficial·
@poetengineer__ Why did you choose Obsidian? I've always imagined that it's node-building animation could somehow be turned into something even more beautiful like this But could never figure out how
English
3
0
23
5.7K
michael_jay
michael_jay@rKaidd·
@dexhorthy i think of it as the "bridge-line formation" that a spider sends out onto a breeze. my job is to bridge between current and the desired (objective) state. the rest is tokens. i suspect in <2 wks "claws" running on mobiles are going to be using camera's IMU's for IRL feedback.
English
0
0
0
23
dex
dex@dexhorthy·
the best ai engineers I know focus on backpressure and verification
English
28
40
632
134.3K
michael_jay
michael_jay@rKaidd·
@andersonbcdefg December’s harness literally deleted itself via a single mistaken escaped cli char. $20 smart switches on the router modern, access points, a fan when 10yr old optiplex overheated after 5hrs of 264 conversions. Oh and turning my tv off to get my attn. what a time
English
0
0
0
659
Ben (no treats)
Ben (no treats)@andersonbcdefg·
curious... you claim to care about "model welfare" and yet you haven't granted your "open claw" the ability to kill itself using the Philips Hue™ Smart Plug... care to elaborate?
Ben (no treats) tweet media
English
53
118
3.6K
154.7K
michael_jay
michael_jay@rKaidd·
why are sota labs 6m behind builders
English
0
0
1
73
Bald Knower ( i cracked ) 🧑🏼‍🦲
Bald Knower ( i cracked ) 🧑🏼‍🦲@BaldKnower·
Buy a Mac mini Download clawdbot Get a basic 3D printer Go in debt if you have to In less than a week I have made enough to buy hundreds of Mac minis with this build
English
542
530
18.2K
2.9M

Keşfet

@sudoingX @ogrizkov @googledevs @Teknium @NousResearch @ErdalToprak @elonmusk @BarackObama