@lennysan Google is crashing, amazon is crashing, coinbase is crashing.... need I go on?
English
wheelofforeplay
621 posts
Engineers don't write code.
PMs are shipping to production.
The design process is dead (there's no time).
Marketing can ship their own campaigns.
SDRs are being replaced by AI.
Everyone's a data scientist now.
What a time to be alive.
English
@AI_with_Eric Will be publishing the most comprehensive set of local AI benchmarks. Not long now.
English
It’s kind of crazy but the shitstorm of supply chain issues has created a new best-in-class local AI deployment: M5 Max MacBook clusters.
- The memory unit economics are great - each MacBook has 128GB @ 614GB/s for $5k
- M5 Max added tensor cores (Apple Neural Accelerators) with 4x compute of M4 Max (~60TFLOPS fp16)
- You can cluster them with RDMA over Thunderbolt 5 and @exolabs for ~linear scaling (so memory bandwidth really is additive: 4 x MacBooks are 512GB @ 2456GB/s)
You can’t get Mac Studios right now, so customers are buying MacBook clusters instead. A government is running this in prod.
This setup is best for low-batch decode-heavy inference (all memory bound) and transcription (super fast and cheap on apple silicon).
Alex Cheema@alexocheema
It is unconventional but it actually works, depending on the workload of course. There are strengths and weaknesses for sure. There are some real deployments (governments, big companies) running this setup in production (pods of 4 MacBooks). It's the best price to performance for many workloads (e.g. transcription, low batch LLM inference). They landed on this themselves as the best hardware to run their workloads on. Can share more in private if you are interested (don't want to turn this into a sales pitch for exo). If Apple actually sold us an M5 Max / M5 Ultra Mac Studio, then we'd use that. But we could be waiting until October for that (or longer, the supply chain issues seem pretty bad). It's the same M5 Max chip in the MacBook as the Mac Studio, and it goes up to 128GB unified memory. Each chip has 614GB/s memory bandwidth (2.24x DGX Spark). I would say the main downside (which we should make more clear) is the software ecosystem - it's still quite immature. It has got much better in the last year e.g. clustering came a long way with low-latency RDMA in macOS 26.2.
English
@alexocheema localmaxxing.com benchmarks or this just theory crafting
English
@jun_song 3090 can be a cheaper alternative to most room heating costs though so keep that in mind if you live anywhere where you often need heating
English
Best budget local llm hardware comparison:
3090 vs Mac Studio M1 Max 64gb
Price : both ~$2k to set up
Pros on 3090 : much better performance (27b vs 35b at similar tok/s)
Pros on Mac : Power efficiency, bigger RAM, zero heat/noise, reliability
English
@jun_song What are peopld using to create read only hermes agents?
English
Frequent question : What is my local llm setup now
Macbook Pro M5 Max 128GB
- running Deepseek-V4-Flash-JANGTQ2
- multi agents set up with Minimax, GLM plan
- Harness : Hermes Agent
2xDGX Spark
- post-training models and world AGI research 24/7
- SuperGemma4-26b agent sub
English
English
English
🚨🗣️ Micah Richards DROPS BOMBSHELL over the VAR controversy involving Arsenal FC
🗣️ Richards: “If this whole VAR situation is properly investigated, then serious questions will need to be asked.”
“You cannot have this level of controversy around key title-deciding moments without scrutiny.”
“And if any wrongdoing or major officiating failures are uncovered, people will start talking about consequences… even potential points deductions.”
😳 “That changes everything.”
“Because now we are no longer talking about one bad decision.”
“We are talking about the integrity of the Premier League title race.”
“I’m not accusing Arsenal of anything directly.”
“But football fans deserve transparency, especially when margins are this small at the top.”
“If the league wants trust from supporters, they must address every controversy properly.” 🔥⚖️
English
I'm giving away five $200 Max plans for @DevinAI
If you're using tools like Claude Code, Codex, or Cursor and haven't tried @DevinAI, comment below with what you're building to be eligible. ⚡️
ellis@DriscollEllis
all you nerds hyperventilating about your claude code config would be gobsmacked by the sophistication devin has had for months. cc is catching up but still relatively unserious
English
@MitoGouken Its the Pro models, the moment the pro model comes out I dont want to upgrade nor do I want to play on my base model
English
- Marathon flopou
- Saros flopou
- GOW Sons of Sparta flopou
- PlayStation 5 tendo uma grande queda de vendas no último relatório financeiro
- Mais de 700 milhões em prejuízo com a Bungie
- Horizon Hunters Gathering é o próximo a flopar
- FairGames provavelmente vai até ser cancelado
PlayStation tá tendo um ano TERRÍVEL até agora.
Português
@BrendanFalk Aussie Engineer here, can you follow me so I can dm you, pretty sure I can smash your test.
Built ainative.school in a few hours
English
If you're an Aussie engineer and think you could crush this interview, please DM me!
We have 100s of thousands of users, millions in revenue, and I would argue the best AI app builder. Lots more to do!
Brendan Falk@BrendanFalk
I believe we've found the best AI-native coding interview We call it the “Composer 1 interview” Candidates get 1 hour to build a real, medium-sized project live The only constraint: they have to use Cursor’s Composer 1 model
English
@mwfowlie @BrendanFalk Sounds like much more reasonable intelligence test to me than the usual shit process most companies put you through
English
@BrendanFalk Why would you use this sort of interview technique? I’m not complaining. I’d be good at it.
But typically you want to measure intelligence or a close proxy to it, not a skill that can be learned quickly.
English
wheelofforeplay retweetledi
🚨Oliver Glasner just delivered one of the coldest press conference responses of the season
Reporter asked him about the title race and whether Palace could influence it…
Glasner replied:
“I checked my pay slips… I didn’t get any money from Arsenal or City.”
Then he finished with THIS:
“Our influence in the title race is definitely less than VAR.”
English
No outrage. No conspiracy theories. No endless debates.
Of course it’s not Arsenal..
English
@Shpeshal_Nick Deploy Saros to PC and Ill buy it right now
English
I swear if Saros doesn’t succeed and we lose Housemarque…
English
Capital gains tax shouldn’t exist.
I risk my money. I build the business. I make the investment. I do the work. I take the risk.
So why the hell should the government take a cut of my success?
They risk nothing. They create nothing. They just take.
Parasites. F’en parasites.
English
@alexocheema Can you benchmark this using localmaxxing.com
English
4 x M5 Max MacBooks clustered with RDMA:
512GB @ 2456GB/s, $20k, 560W, quiet.
Find me a better deal, that I can buy today.
Ivan Kuleshov@Merocle
Cluster testing is currently in progress
English
@jun_song @dealignai 128gb is absolutely worth it, deepseek v4 flash is significantly better than Qwen3.6
English
Local LLM fits for Mac RAM size (5/14):
>~32gb : SuperGemma4-e4b-mlx
>32~64gb : Qwen3.6-35b-mlx-6bit
>96~128gb : Minimax-M2.7 / Deepseek-V4-Flash (JANGTQ by @dealignai)
>256gb : Xiaomi-MiMo-V2.5
>512gb : GLM-5.1-RAM-420GB-MLX
Follow and keep on updates
English
@Edouardmazza @jun_song @dealignai I ran it through multiple coding tasks and I simply found the output of 27B superior to 35B.
35B is definitely faster though, so its a quality (27B) vs speed (35B) choice.
English
@jun_song @dealignai Fair, my only experience is not with a mac
English