XOutlier
Download .zipAdd to Chrome

Alex Cheema

@alexocheema55k followers

building @exolabs | prev @UniOfOxford We're hiring:

Open on X
Top outlier
328×
Best organic post
Peak views
3.5M
Single post
Bookmarks
77k
Across 89 organic posts
Last scan
Oct 16
Community cache
Posts 89 posts
#DatePost↓ OutlierViewsLikesRepostsBookmarks
1Nov 9, '24M4 Mac Mini AI Cluster Uses @exolabs with Thunderbolt 5 interconnect (80Gbps) to run LLMs distributed across 4 M4 Pro Mac Minis. The cluster is small (iPhone for reference). It’s r328×3.5M25k2.5k11k
2Jan 28Running Kimi K2.5 on my desk. Runs at 24 tok/sec with 2 x 512GB M3 Ultra Mac Studios connected with Thunderbolt 5 (RDMA) using @exolabs / MLX backend. Yes, it can run clawdbot.310×3.3M8.8k6485k
3Jan 21, '25AGI at home Running DeepSeek R1 across my 7 M4 Pro Mac Minis and 1 M4 Max MacBook Pro. Total unified memory = 496GB. Uses @exolabs distributed inference with 4-bit quantization. Ne181×1.9M10k1k4.9k
4Jan 20, '25I will run AGI at home or die trying. DeepSeek R1 should run fast on these macs. They have a total of 896GB unified memory @ 3557GB/s132×1.4M9.2k5292.6k
5Jul 24, '242 MacBooks is all you need. Llama 3.1 405B running distributed across 2 MacBooks using @exolabs_ home AI cluster120×1.3M5.2k6853.5k
6Jan 27, '25Market close: $NVDA: -16.91% | $AAPL: +3.21% Why is DeepSeek great for Apple? Here's a breakdown of the chips that can run DeepSeek V3 and R1 on the market now: NVIDIA H100: 80GB @117×1.2M7k1k4.4k
7Mar 11, '25Apple have given me early access to 2 maxed out M3 Ultra 512GB Mac Studios ahead of the public release. I will run the full DeepSeek R1 (8-bit) using @exolabs or die trying. The 1T113×1.2M5k2681.1k
8Nov 12, '24Backdoor attempt on @exolabs through an innocent looking PR. Read every line of code. Stay safu.109×1.2M5.5k4221.5k
9Jan 7, '25While Apple has been positioning M4 chips for local AI inference with their unified memory architecture, NVIDIA just undercut them massively. Stacking Project Digits personal compu102×1.1M4.1k5581.8k
10Mar 12, '25Running DeepSeek R1 on my desk Uses @exolabs with Thunderbolt 5 interconnect (80Gbps) to run the full (671B, 8-bit) DeepSeek R1 distributed across 2 M3 Ultra 512GB Mac Studios (1TB92.7×993k6.9k6602.7k
11Sep 6, '25Called it last year. Junior roles down 23%. Senior roles up 14%. Harvard study tracked 285,000 firms. Results: Before AI: 1 senior + 3 juniors = 4 person team After AI: 1 senior + 90.6×970k2.8k3011.4k
12Nov 23, '24Price/performance ratio for AI increased dramatically with the new M4 Mac Mini. LPDDR5 @ 256GB/s with AMD/NVIDIA Linux alternatives will be on the market soon with a similar form f73.5×788k2.9k2461.5k
13Jan 15, '25"Somebody got one of the small versions of Llama to run on Windows 98...We could've been talking to our computers in English for the last 30 years" - @pmarca It was me! I got Llama71.5×766k6.1k4491.7k
14Jun 13, '24Connecting a bunch of iPhones, iPads and MacBooks together over a local network to make one big GPU. Uses Apple’s open source ML library, MLX70.6×756k5.2k5592.2k
15Dec 27, '24Had to stack up 8 Mac Minis to get it running. ~5 tok/sec for now. First time running inference on 8 Mac Minis - performance can be improved a lot (theoretical limit is >10 tok/65.2×698k2.3k179751
16Jun 28, '25fml linkedin is unusable64.3×689k6k2271k
17Dec 26, '24I will run Deepseek-V3-Base 685B on M4 Mac Minis or die trying. 685B MoE with 256 experts -- perfect for Apple Silicon since they have a lot of GPU memory and only a small subset o63.5×681k4.4k2631.6k
18Sep 20, '25If you’re a talented engineer affected by the H-1B changes, come build with us in London @exolabs - SF-level comp (270K-360K base + equity) - Best talent from Europe - Hardcore bui57.2×612k4.9k3091.6k
19Dec 28, '2435.9 tok/sec on Windows 98 🤯 This is a 260K LLM with Llama-architecture. We also tried out larger models. Results in the blog post.53.2×570k2.6k299894
20Mar 5, '25Apple's timing could not be better with this. The M3 Ultra 512GB Mac Studio fits perfectly with massive sparse MoEs like DeepSeek V3/R1. 2 M3 Ultra 512GB Mac Studios with @exolabs 51.1×547k2.7k2981.6k
21Nov 8, '24How much faster is the new MacBook Pro for AI inference? M4 Max is 27% faster with 72 tok/sec compared to 56 tok/sec of the M3 Max with MLX running Gemma 2 9B (4bit). The 27% speed49.4×530k2.4k198773
22Aug 23, '24Most people don't understand @cursor_ai. They think it will enable anyone to ship production code. They think it will replace developers. It won't. It will augment the best develop48.7×522k56355302
23Nov 12, '24M4 Mac AI Coding Cluster Uses @exolabs to run LLMs (here Qwen 2.5 Coder 32B at 18 tok/sec) distributed across 4 M4 Mac Minis (Thunderbolt 5 80Gbps) and a MacBook Pro M4 Max. Local 48.4×518k3.8k3452k
24Dec 25, '24Raspberry Pi 5 is a bit smaller than NVIDIA Jetson Orin Nano Super44.3×475k5.2k199797
26Mar 18, '25NVIDIA DGX Spark. World’s smallest AI supercomputer with 128GB memory29.2×312k1.8k151313
27Apr 5, '25Meta 🤝 Apple Llama 4 + Apple Silicon is a match made in heaven. Here's why: Like DeepSeek V3/R1, all of the new Llama 4 variants are massive sparse MoE models. They have a massive28.8×308k2.6k3271.5k
28Mar 28, '25Is it just me who has switched entirely to Gemini 2.5 Pro? It’s free, fast, smart and has 1M context window. I just have one big chat with 10K+ lines of code and thoughts. It can s28×300k2.4k75818
29Apr 18, '24Llama 3 running locally on iPhone with MLX Built by @exolabs_ team @mo_baioumy h/t @awnihannun MLX & @Prince_Canuma for the port27.6×296k910137463
30Oct 15, '25NVIDIA sent us 2 DGX Sparks. For a while we wondered what we would do with them. The memory bandwidth is 273GB/s making it 3x slower than an M3 Ultra (819GB/s) for batch_size=1 inf26.3×281k1.4k123608
31Dec 16, '24"It's on my todo to build my own stack of apple minis and run Llama on my own little cluster on my desk." - @garrytan Quick AI homelab tutorial: - Install @exolabs on each mini (op22.9×245k1.6k1621.4k
32Jul 26, '24Llama 3.1 70b beamed to my iPhone from my @exolabs_ home AI cluster of 2 MacBooks and 1 Mac Studio My own private GPT-4 assistant at home / on the go22.7×244k1.4k147705
33Sep 13, '24The future of AI is open source and decentralized.22.6×242k85158108
34Jun 14, '24Brought my Apple device AI cluster with me on my 7 hr flight @exolabs_ iPhones, iPad and MacBook connect together on local network to make one big GPU. 1 device = run small models 21.2×227k1.5k131637
35Jul 1, '24🥇 I won the @__tinygrad__ x @AGIHouseSF hackathon solo (+support of @realGeorgeHotz) Hacked an @apple Mac Mini distributed AI training cluster with thunderbolt-4 interconnect (40G18.9×202k2.1k133562
36Aug 26, '24Managed to run Llama 3.1 405b on my flight. Uses @exolabs_ to distribute the AI model across 2 MacBooks. GPT-4o-level model running offline, so I can play games, ask questions and 18.7×201k1.1k62495
37Jul 15, '24Running Llama-3-70B at home with @exolabs_ Combines the compute of all these devices to make one big GPU: - iPhone 15 Pro Max - iPad Pro M4 - Galaxy S24 Ultra - MacBook Pro M2 and 18.4×198k1.2k162938
38Dec 7, '24Preparing to run the new Llama 3.3 70B on my mac cluster. Downloading model shards onto 3 x M4 Pro Mac Mini and 1 x M3 Max MacBook Pro. AI cluster is connected by Gigabit ethernet 17×182k1.3k114605
39Aug 24, '24Got 2 MacBooks with me running distributed Llama-3.1-405b on @exolabs_ ready for my flight. Portable intelligence in the air, works fully offline with Thunderbolt 4 interconnect.16.8×180k1.5k123639
40Aug 31, '24On the power efficiency of Apple M chips. @exolabs_ with the MLX backend can get ~100% GPU utilisation out of each MacBook Pro M3 Max. The GPU has 128GB of memory (bandwidth ~400GB15.7×168k785108444
42Jul 23, '24This is the biggest @exolabs_ cluster I've seen so far. 23 x Mac Mini 16GB (total 368GB). Enough to run Llama-3-405b at Q4_K_S (230GB w/o cache).14.2×152k1.4k127514
43Aug 5, '24Someone saw the mess we had with @exolabs_ getting hold of the ~1TB of Llama 3.1 405b weights so they started a company: Physical drives shipped to your door full of the latest LLM13.7×147k81376257
44Aug 19, '24an e-reader with an LLM running on it (not a phone) @exolabs_ will make all underutilised hardware useful again12.8×137k1k84458
46Aug 31, '2423 iMacs running distributed llama 3 405b with @exolabs_11.3×121k82369121
47Oct 30, '24For the first time in history, you can run Llama 405B at reading speed on your own hardware with this setup using @exolabs and M4 Max. M4 Max has 546GB/s of memory bandwidth. The t11.1×119k60658300
48Mar 30, '25That was me! I got Llama running on a Pentium II machine with 128MB RAM running Windows 98. Details below.10.9×117k1.1k82218
49Aug 23, '24Each device in the @exolabs_ mesh network downloads a shard of the AI model. Each shard “sticks” to a device, so you don’t need to re-download parts of the model. You can add as ma10.9×117k58360267
50Mar 17, '25how @exolabs v2 will enable 128 mac mini clusters (unironically): - search-based orchestration of ML jobs across arbitrary distributed device topologies (bitter lesson) - low inter10.7×115k90495272
51Aug 21, '24Connected my MacBook Pro GPU to my Linux laptop NVIDIA GPU using @exolabs_ Running really large AI models with 193 TFLOPs of compute, combining the M3 GPU + RTX 409010.6×113k57365301
52Jul 19, '24Speed running my home AI cluster running distributed inference across 2 MacBooks and 2 Mac Minis. @exolabs_ displays a real-time network topology as devices discover each other ove10.3×110k1.1k106703
53Jul 21, '24You don't need a H100 to run Llama-3-405b. 2 MacBooks and 1 Mac Studio will do the job, with @exolabs_ to aggregate the memory/compute. I'm ready for you, Llama-3-405b.9.7×104k958106479
54Jun 19, '25"Mac Minis for example are a very good fit" - @karpathy @karpathy shouted out my work on @exolabs in his keynote at @ycombinator AI SUS! Here's the breakdown: Right now most AI wor9.7×104k40643189
55Jun 18, '25.@karpathy shouted out my work on @exolabs at @ycombinator AI SUS! “we use LLMs similarly to mainframes in the ‘70s - compute is timeshared by having a slice in the batch dimension9.4×101k54236138
56Jul 18, '24How long does it take to get distributed inference running locally across 2 MacBook GPUs from a fresh install? About 60 seconds, running @exolabs_ Watch till the end, I chat to the9.3×99k58045431
58Dec 29, '24llama2.c in pure C from @karpathy was the only OSS project I could find that would enable running an LLM on the Windows 98 Pentium II machine. I couldn't get modern cpp (e.g. llama8.5×91k95937190
59Jan 4, '25Most of the compute in the world is not in AI data centers. It's distributed around the globe. It's dark compute. iPhones are dark compute. In 2024, Apple sold over 200M iPhones. I8.2×88k42450135
60Jul 5, '24You thought @karpathy was kidding about all those Mac Minis?8.1×87k3451466
61Oct 20, '24The future is now. Microsoft have open sourced their inference framework for ternary models (aka 1-bit LLMs)7.6×81k49945285
62Jan 27, '25Latest @fireship_dev video featured my run of DeepSeek R1 on M4 Mac Minis. Apple Silicon dominates in memory/bw unit economics, ideal for huge MoE models like R1 at batch_size=1 (t7.5×80k42638176
64Dec 25, '24Merry EXmas7.1×76k2k4470
65Jan 2, '25A few people asked me how I made these animations. They don't believe me when I tell them I used Cursor with Claude Sonnet. These are raw text .svg files. Sonnet is surprisingly go7.0×75k70155645
66Dec 8, '24M4 Mac Mini AI cluster running Llama 3.3 70B Preliminary results show Gigabit ethernet switch (4 tok/sec) is slower than Thunderbolt-5 with @exolabs Benchmarks including Thunderbol6.8×73k49155163
67Jul 10, '24For those asking for the @__tinygrad__ Mac Mini Cluster tutorial, I’m cooking something up for you now… @karpathy @exolabs_6.7×72k1k46350
68Apr 24, '24Phi-3 running locally on an iPhone using MLX Fully offline, it’s fast! Credit to @exolabs_ team @mo_baioumy h/t @awnihannun for the speedy model impl in MLX6.6×71k33936191
69Oct 15, '243 Reasons Why @__tinygrad__ Will Win The ML Library War. 1. tinygrad searches its kernels. tinygrad uses BEAM search and soon MCTS to search for optimal kernels. You only have to w6.3×68k36937245
70Mar 14, '25Now imagine this running distributed over Macs. @exolabs v2 operates at the level of abstraction of the @__tinygrad__ UOp graph so we'll have distributed cryptography out of the bo5.7×61k3111865
71Nov 24, '247 x M4 Mac Mini Cluster The next frontier will be double-digit M4 clusters, then triple-digit. Releasing benchmarks in the open with @exolabs soonTM.5.7×61k53639115
72Jun 25, '24Linear scaling achieved internally @exolabs_ @karpathy5.5×59k6613297
73Apr 18, '25You may not like it, but this is what peak performance looks like.5.4×58k9145985
74Feb 21, '258 months ago, exo was a hackathon project. Today it's front page of The Wall Street Journal @WSJ. We're a real company now (I guess..?), we raised some money from a few investors l5.3×57k68956119
77Jul 23, '24Anyone in the Bay Area with fast download or weights for Llama-3.1-405b? cc @realGeorgeHotz @AGIHouseSF Faster to drive to you and transfer the weights physically OTC Final thing n4.9×53k3241672
78Sep 9, '24Yes, exo will support Apple TV's.4.5×49k3471643
79Oct 16, '24Running @nvidia Nemotron-70B distributed across 2 MacBooks. Uses @exolabs with the MLX backend.4.3×47k38325170
80Jan 30, '25Other people are starting to run DeepSeek R1 on their own hardware. Tutorial coming soonTM4.1×44k58851182
81Jul 2, '24Following the Mac Mini AI cluster hack win, I got to visit Apple in Cupertino with the @exolabs_ team4.0×43k3181036
82Aug 19, '24Your own multimodal AI assistant at home Running on an iPhone and 2 MacBooks using @exolabs_ to aggregate the compute4.0×42k33640153
83Dec 31, '24Now when I go out for a walk, I take the Nokia with me. Distraction free (I don’t get notifications / can’t see my feed) but I can still talk to an LLM and it has access to almost 3.8×41k40931129
85Oct 29, '24Training AI models on Apple Silicon just got 3x better.3.4×36k3542149
86Jun 3, '25Running DeepSeek R1 0528 on the tube 1TB of portable high-speed unified memory in a box3.3×36k3872459
87Nov 11, '24exo is trending on GitHub again. exo blew up way sooner than I expected. When we open sourced it 119 days ago, I was planning for a slow start where we would work in the open, pick3.2×35k40234119
88Nov 8, '24Happy M4 day! AI benchmarks coming shortly.3.2×34k327715
89Dec 7, '24Running @exolabs AI cluster with @Raspberry_Pi 4. exo leverages any spare hardware you have available and puts it to use on AI workloads by splitting them across all devices, You k3.1×33k39033101
90Apr 10, '25This is 10x easier in the new @exolabs - one click to install the exo background service + auto discovery + topology-aware heterogeneous device compute plans. MLX is awesome tech. 3.0×32k2903877
91Dec 31, '24Being a founder is like getting punched in the face 99 times and still hoping the 100th time is a kiss instead2.9×31k3511134
92Feb 17, '25AI homelab on 5 M2 Ultra Mac Studios using @exolabs Imagine this with M4 Ultras. They'll likely use Apple UltraFusion to fuse together two M4 Max dies, for a massive 256GB of unifi2.9×31k3312982
94Apr 30, '25               2.9×31k4692039
95Dec 25, '24Adding benchmarks for DeepSeek V3-600B on M4 Mac Mini Cluster2.6×28k3951967
96Dec 30, '24Linear scaling with 5 nodes over Gigabit ethernet. We can push this a lot further with Thunderbolt-5 to 100+ macs. Ethernet is a lot easier from a network engineering perspective s2.0×21k3002543
98Mar 11, '25Last night in London I got a call about an exciting opportunity in SF. So exciting that I booked the first flight out to SF. I'm jetlagged and caffeinated, but super hyped for this1.8×19k293214
99Jul 22, '24Real-time distributed inference monitoring is live on @exolabs_ home AI cluster It comes with out of the box support for @PrometheusIO metrics and @grafana dashboards Helps diagnos1.7×18k29334108

Scan another creator

@
@karpathy@levelsio@paulg@naval@marclou@illyism

Run this on any account.

XOutlier builds this table for every public profile you open on x.com, with live data and every post, not just the top 50.

Add to Chrome Download .zip