This year, we\u2019ve pledged to buy $20K worth of compute so we can experiment with AI doing science, agent orchestration, and maybe even use it as edge compute to mimic scale for our apps and test how production-ready they are.
We already have a DGX Spark and two RTX 5080s, and will eventually connect all GPUs over Exo Labs to be able to run upwards of 400B parameter models locally at hopefully adequate inference throughput.
We decided to run a slightly absurd experiment on the Spark.
Inspired by The Hitchhiker\u2019s Guide to the Galaxy, where a supercomputer spends millions of years searching for the meaning of life, we tried to recreate a tiny, ridiculous version of that.
We\u2019re running a Gemma 4 4B (at FP8 precision) parameter model to generate plausible \u201cmeaning of life\u201d questions from different angles, and a Gemma 4 26B MoE model (at FP4 precision) to reason through them, judge them, and score the answers.
Basically: what do LLMs think the meaning of life is when you let them think for 24 hours?
This is the kind of experiment that feels stupid to run on API consumption, but run moment the compute is yours.
Right now the winning answer is "The meaning of life is to utilize the emergent capacity for consciousness to assign value to a universe that is otherwise driven by blind physical laws."
We\u2019re already 2 hour in. The experiment will run for another 22 hours, and you can track it live below.
Live website: https://eternal-question.vercel.app
GitHub repo: https://github.com/piyussh01/eternalquestion"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"We're spending 24 hours using local LLMs to search for the meaning of life"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://eternal-question.vercel.app/"}},"_tags":["comment","author_piyussh","story_48473062"],"author":"piyussh","comment_text":"At our SF hacker house, The Garden, we believe compute is the new commodity to own.
This year, we\u2019ve pledged to buy $20K worth of compute so we can experiment with AI doing science, agent orchestration, and maybe even use it as edge compute to mimic scale for our apps and test how production-ready they are.
We already have a DGX Spark and two RTX 5080s, and will eventually connect all GPUs over Exo Labs to be able to run upwards of 400B parameter models locally at hopefully adequate inference throughput.
We decided to run a slightly absurd experiment on the Spark.
Inspired by The Hitchhiker\u2019s Guide to the Galaxy, where a supercomputer spends millions of years searching for the meaning of life, we tried to recreate a tiny, ridiculous version of that.
We\u2019re running a Gemma 4 4B (at FP8 precision) parameter model to generate plausible \u201cmeaning of life\u201d questions from different angles, and a Gemma 4 26B MoE model (at FP4 precision) to reason through them, judge them, and score the answers.
Basically: what do LLMs think the meaning of life is when you let them think for 24 hours?
This is the kind of experiment that feels stupid to run on API consumption, but run moment the compute is yours.
Right now the winning answer is "The meaning of life is to utilize the emergent capacity for consciousness to assign value to a universe that is otherwise driven by blind physical laws."
We\u2019re already 2 hour in. The experiment will run for another 22 hours, and you can track it live below.
Live website: https://eternal-question.vercel.app
GitHub repo: https://github.com/piyussh01/eternalquestion","created_at":"2026-06-10T08:07:58Z","created_at_i":1781078878,"objectID":"48473063","parent_id":48473062,"story_id":48473062,"story_title":"We're spending 24 hours using local LLMs to search for the meaning of life","story_url":"https://eternal-question.vercel.app/","updated_at":"2026-06-10T10:22:23Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"storus"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"Does this also run with Exo Labs' token pre-fill acceleration using DGX Spark? I.e. take 2 Sparks and 2 MacStudios and get a comparable inference speed to what 2x M5 Ultras will be able to do?"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Kimi K2 1T model runs on 2 512GB M3 Ultras"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://twitter.com/awnihannun/status/1943723599971443134"}},"_tags":["comment","author_storus","story_46262734"],"author":"storus","comment_text":"Does this also run with Exo Labs' token pre-fill acceleration using DGX Spark? I.e. take 2 Sparks and 2 MacStudios and get a comparable inference speed to what 2x M5 Ultras will be able to do?","created_at":"2025-12-14T19:55:38Z","created_at_i":1765742138,"objectID":"46266222","parent_id":46262734,"story_id":46262734,"story_title":"Kimi K2 1T model runs on 2 512GB M3 Ultras","story_url":"https://twitter.com/awnihannun/status/1943723599971443134","updated_at":"2026-03-05T23:11:53Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"aimanbenbaha"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"Exo-Labs is an open source project that allows this too, pipeline parallelism I mean not the latter, and it's device agnostic meaning you can daisy-chain anything you have that has memory and the implementation will intelligently shard model layers across them, though its slow but scales linearly with concurrent requests.
Exo-Labs: https://github.com/exo-explore/exo"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"macOS 26.2 enables fast AI clusters with RDMA over Thunderbolt"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://developer.apple.com/documentation/macos-release-notes/macos-26_2-release-notes#RDMA-over-Thunderbolt"}},"_tags":["comment","author_aimanbenbaha","story_46248644"],"author":"aimanbenbaha","comment_text":"Exo-Labs is an open source project that allows this too, pipeline parallelism I mean not the latter, and it's device agnostic meaning you can daisy-chain anything you have that has memory and the implementation will intelligently shard model layers across them, though its slow but scales linearly with concurrent requests.
Exo-Labs: https://github.com/exo-explore/exo","created_at":"2025-12-13T13:25:13Z","created_at_i":1765632313,"objectID":"46254356","parent_id":46249754,"story_id":46248644,"story_title":"macOS 26.2 enables fast AI clusters with RDMA over Thunderbolt","story_url":"https://developer.apple.com/documentation/macos-release-notes/macos-26_2-release-notes#RDMA-over-Thunderbolt","updated_at":"2026-03-05T23:11:06Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"storus"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"I run it all the time, token generation is pretty good. Just large contexts are slow but you can hook a DGX Spark via Exo Labs stack and outsource token prefill to it. Upcoming M5 Ultra should be faster than Spark in token prefill as well."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"DeepSeek-v3.2: Pushing the frontier of open large language models [pdf]"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/main/assets/paper.pdf"}},"_tags":["comment","author_storus","story_46108780"],"author":"storus","children":[46113264],"comment_text":"I run it all the time, token generation is pretty good. Just large contexts are slow but you can hook a DGX Spark via Exo Labs stack and outsource token prefill to it. Upcoming M5 Ultra should be faster than Spark in token prefill as well.","created_at":"2025-12-01T20:03:13Z","created_at_i":1764619393,"objectID":"46112394","parent_id":46112104,"story_id":46108780,"story_title":"DeepSeek-v3.2: Pushing the frontier of open large language models [pdf]","story_url":"https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/main/assets/paper.pdf","updated_at":"2026-03-05T23:04:54Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"adav"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"EXO Labs are doing interesting things with LLM clusters on commodity hardware. I'd quite like to plug my MacBook next to my iPad and get the extra local AI oomph."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"iPad Pro with M5 chip"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.apple.com/newsroom/2025/10/apple-introduces-the-powerful-new-ipad-pro-with-the-m5-chip/"}},"_tags":["comment","author_adav","story_45591905"],"author":"adav","comment_text":"EXO Labs are doing interesting things with LLM clusters on commodity hardware. I'd quite like to plug my MacBook next to my iPad and get the extra local AI oomph.","created_at":"2025-10-15T21:40:50Z","created_at_i":1760564450,"objectID":"45598664","parent_id":45596495,"story_id":45591905,"story_title":"iPad Pro with M5 chip","story_url":"https://www.apple.com/newsroom/2025/10/apple-introduces-the-powerful-new-ipad-pro-with-the-m5-chip/","updated_at":"2026-03-05T22:53:40Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"adwn"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"> EXO Labs showed otherwise by getting a 300K-parameter LLM to run on a Pentium II with only 128 MB of RAM at about 50 tokens per second
50 token/s is completely useless if the tokens themselves are useless. Just look at the "story" generated by the model presented in your link: Each individual sentence is somewhat grammatically correct, but they have next to nothing to do with each other, they make absolutely no sense. Take this, for example:
"I lost my broken broke in my cold rock. It is okay, you can't."
Good luck tuning this for turn-based conversations, let alone for solving any practical task. This model is so restricted that you couldn't even benchmark its performance, because it wouldn't be able to follow the simplest of instructions."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Cray versus Raspberry Pi"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.aardvark.co.nz/daily/2025/0611.shtml"}},"_tags":["comment","author_adwn","story_44244595"],"author":"adwn","children":[44281927],"comment_text":"> EXO Labs showed otherwise by getting a 300K-parameter LLM to run on a Pentium II with only 128 MB of RAM at about 50 tokens per second
50 token/s is completely useless if the tokens themselves are useless. Just look at the "story" generated by the model presented in your link: Each individual sentence is somewhat grammatically correct, but they have next to nothing to do with each other, they make absolutely no sense. Take this, for example:
"I lost my broken broke in my cold rock. It is okay, you can't."
Good luck tuning this for turn-based conversations, let alone for solving any practical task. This model is so restricted that you couldn't even benchmark its performance, because it wouldn't be able to follow the simplest of instructions.","created_at":"2025-06-15T10:50:05Z","created_at_i":1749984605,"objectID":"44281623","parent_id":44281285,"story_id":44244595,"story_title":"Cray versus Raspberry Pi","story_url":"https://www.aardvark.co.nz/daily/2025/0611.shtml","updated_at":"2025-06-17T03:09:36Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rahen"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"I would have thought the same, but EXO Labs showed otherwise by getting a 300K-parameter LLM to run on a Pentium II with only 128 MB of RAM at about 50 tokens per second. The X-MP was in the same ballpark, with the added benefit of native vector processing (not just some extension bolted onto a scalar CPU) which performs very well on matmul.
https://www.tomshardware.com/tech-industry/artificial-intell...
John Carmack was also hinting at this: we might have had AI decades earlier, obviously not large GPT-4 models but useful language reasoning at a small scale was possible. The hardware wasn't that far off. The software and incentives were.
https://x.com/ID_AA_Carmack/status/1911872001507016826"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Cray versus Raspberry Pi"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.aardvark.co.nz/daily/2025/0611.shtml"}},"_tags":["comment","author_rahen","story_44244595"],"author":"rahen","children":[44281623],"comment_text":"I would have thought the same, but EXO Labs showed otherwise by getting a 300K-parameter LLM to run on a Pentium II with only 128 MB of RAM at about 50 tokens per second. The X-MP was in the same ballpark, with the added benefit of native vector processing (not just some extension bolted onto a scalar CPU) which performs very well on matmul.
https://www.tomshardware.com/tech-industry/artificial-intell...
John Carmack was also hinting at this: we might have had AI decades earlier, obviously not large GPT-4 models but useful language reasoning at a small scale was possible. The hardware wasn't that far off. The software and incentives were.
https://x.com/ID_AA_Carmack/status/1911872001507016826","created_at":"2025-06-15T08:43:25Z","created_at_i":1749977005,"objectID":"44281285","parent_id":44280951,"story_id":44244595,"story_title":"Cray versus Raspberry Pi","story_url":"https://www.aardvark.co.nz/daily/2025/0611.shtml","updated_at":"2025-06-17T13:17:08Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ryzvonusef"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"Alex Cheema is the guy behind the EXO Labs, and he got the two MS Ultra from Apple early (iirc he paid full price, he just got to skip the line).
Local AI is become increasingly viable, I expect a trend towards local 'mainframe' to resurge again; a tiny rack with like ~$50K of equipment that replaces at least a dozen paper-processing staff would be a viable expense for any mid-sized company."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"\"11 tok/SEC running 671B DeepSeek R1 on two M3 Ultra\" \u2013 Alex Cheema"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://twitter.com/alexocheema/status/1899735281781411907"}},"_tags":["comment","author_ryzvonusef","story_43341684"],"author":"ryzvonusef","comment_text":"Alex Cheema is the guy behind the EXO Labs, and he got the two MS Ultra from Apple early (iirc he paid full price, he just got to skip the line).
Local AI is become increasingly viable, I expect a trend towards local 'mainframe' to resurge again; a tiny rack with like ~$50K of equipment that replaces at least a dozen paper-processing staff would be a viable expense for any mid-sized company.","created_at":"2025-03-12T10:38:34Z","created_at_i":1741775914,"objectID":"43341743","parent_id":43341685,"story_id":43341684,"story_title":"\"11 tok/SEC running 671B DeepSeek R1 on two M3 Ultra\" \u2013 Alex Cheema","story_url":"https://twitter.com/alexocheema/status/1899735281781411907","updated_at":"2025-03-18T19:47:51Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"dtnewman"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"Check out exo labs blog. https://blog.exolabs.net/day-2/"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Ask HN: Deepseek v3 how can I check on local computer?"}},"_tags":["comment","author_dtnewman","story_42532029"],"author":"dtnewman","comment_text":"Check out exo labs blog. https://blog.exolabs.net/day-2/","created_at":"2024-12-29T15:20:18Z","created_at_i":1735485618,"objectID":"42540435","parent_id":42532029,"story_id":42532029,"story_title":"Ask HN: Deepseek v3 how can I check on local computer?","updated_at":"2024-12-29T15:23:09Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"fsiefken"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["exo","labs"],"value":"It's good, but I'm already paying for GPT4o and Sonnet.\nHow much memory does this need? If Alex Cheema (Exo Labs, Oxford) https://x.com/ac_crypto/status/1815969489990869369 could run Llama 3.1 405 Model on 2 macbooks, does this mean this can run on one macbook?"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Hermes 3: The First Fine-Tuned Llama 3.1 405B Model"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://lambdalabs.com/blog/unveiling-hermes-3-the-first-fine-tuned-llama-3.1-405b-model-is-on-lambdas-cloud"}},"_tags":["comment","author_fsiefken","story_41260040"],"author":"fsiefken","children":[41261559,41265365],"comment_text":"It's good, but I'm already paying for GPT4o and Sonnet.\nHow much memory does this need? If Alex Cheema (Exo Labs, Oxford) https://x.com/ac_crypto/status/1815969489990869369 could run Llama 3.1 405 Model on 2 macbooks, does this mean this can run on one macbook?","created_at":"2024-08-15T22:15:09Z","created_at_i":1723760109,"objectID":"41261028","parent_id":41260040,"story_id":41260040,"story_title":"Hermes 3: The First Fine-Tuned Llama 3.1 405B Model","story_url":"https://lambdalabs.com/blog/unveiling-hermes-3-the-first-fine-tuned-llama-3.1-405b-model-is-on-lambdas-cloud","updated_at":"2024-09-20T17:39:42Z"}],"hitsPerPage":20,"nbHits":10,"nbPages":1,"page":0,"params":"query=%22exo+labs%22&hitsPerPage=20&advancedSyntax=true&analyticsTags=backend","processingTimeMS":9,"processingTimingsMS":{"_request":{"queue":24,"roundTrip":16},"afterFetch":{"merge":{"entries":{"decompress":1,"total":1},"mergeLoop":{"prepareNextHit":6,"total":6},"total":8},"total":8},"total":9},"query":"\"exo labs\"","serverTimeMS":34}