Top all-time · #8512
Sam McLeod
Personal blog · smcleod.net · on HN as smcleodprogrammingsystemsai
Sam McLeod's personal blog at smcleod.net covers programming, systems and ai. Since 2015, 9 different Hacker News readers have linked it 36 times, most often “Bringing K/V context quantisation to Ollama”. It is still writing: the last post was on 2026-10-10, and you can follow it by feed.
Too few sampled comments (5) to say how often HN recommends it.?
Why HN links it
Commenters link his posts and tools on running local LLMs: KV cache quantisation, a VRAM estimator, and mirroring GitLab to GitHub. The comments cite them as references rather than judging them.
Summarised by Claude from a sample of 5 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
The personal blog of Sam McLeod.
Most-linked posts?
- Bringing K/V context quantisation to Ollama3 people · 3 mentions · first linked 2024-12-05
- Building a high performance SSD SAN, part 12 people · 3 mentions · first linked 2015-03-15
- /2 people · 3 mentions · first linked 2016-11-10
What Hacker News says?
I agree with those that say RSS never went away, for the last 10 years or so I’ve noticed my dependence on RSS grow and it’s critical to my keeping track of news across a number of blogs and websites. I wrote a shitty little blog post on it last year: https://smcleod.net/thoughts/2017/09/22/return-of-the-rss.ht...
Comment linking /thoughts/2017/09/22/return-of-the-rss.html
on “It's time to head back to RSS?”Submitted Bringing K/V context quantisation to Ollama as a story · 220 points
HN discussionI very much agree with your sentiment here. I tried to encapsulate that to some degree when writing something (perhaps poorly?) recently actually - https://smcleod.net/2025/03/the-democratisation-paradox-what...
Comment linking /2025/03/the-democratisation-paradox-what-history-teaches-us-about-ai/
on “Thoughts on thinking”Submitted Building a high performance SSD SAN, part 1 as a story · 69 points
HN discussionI truly love RSS, combining Feedly with Reeder makes for a fantastic experience. I wrote a brief post on how I use RSS recently: https://smcleod.net/thoughts/return-of-the-rss/
Comment linking /thoughts/return-of-the-rss/
on “RSS: there's nothing better”Submitted My recommended Firefox addons for 2022 as a story · 20 points
HN discussionThank you for your kind words. I'm very glad to hear our quality is improving and you're having pain-free upgrades. We're proud that GitLab allows everyone to contribute, even organizations without a large budget https://about.gitlab.com/strategy/ Thanks for using GitLab and writing about it https://smcleod.net/mirror-gitlab-to-github/
Comment linking /mirror-gitlab-to-github/
on “GitLab 8.5 released”Submitted Stop Trying to Hire with Titles Like ‘DevOps Engineer’ or ‘Cloud Engineer’ as a story · 5 points
HN discussionYou can do pretty amazing things with well designed (onprem) networked storage with NVMe drives/arrays. I replaced the company I was working with at the time's traditional "enterprise" HPE SANs with standard linux servers running a mix of NVMe and SATA SSDs that provided highly available, low latency and decent throughput iSCSI via network. Gen 1 back in 2014/2015 did something like 70K random 4k read/write IOP/s pe…
Comment linking /tech/2015/07/24/scsi-benchmarking/
on “Modern storage is plenty fast, but the APIs are bad”Submitted The Aeroplane Flies High, Turns Left, Looks Right and Might Reach Its Destination as a story · 3 points
HN discussionEven at q8_0? I thought it wasn't bad just like the models itself. But very interested to hear. And q8_0 already halves the memory usage compared to fp16. One of the ollama Devs called the quality impact negligible at q8_0: https://smcleod.net/2024/12/bringing-k/v-context-quantisatio... But perhaps quantifying the KV cache does not scale as gracefully as the model itself?
Comment linking /2024/12/bringing-k/v-context-quantisation-to-ollama/
on “Building a personal, private AI computer on a budget”Submitted New Apple Silicon M4 and M5 HiDPI Limitation on 4K External Displays as a story · 1 points
HN discussionSorry this isn't more obvious. Ideally VRAM usage for the context window (the KV cache) becomes dynamic, starting small and growing with token usage, whereas right now Ollama defaults to a size of 2K which can be overridden at runtime. A great example of this is vLLM's PagedAttention implementation [1] or Microsoft's vAttention [2] which is CUDA-specific (and there are quite a few others). 1M tokens will definitely…
Comment linking /2024/12/bringing-k/v-context-quantisation-to-ollama/
on “Qwen2.5-1M: Deploy your own Qwen with context length up to 1M tokens”Submitted MBA Washing as a story · 1 points
HN discussionWay back when, in physical land - I used STONITH for that! https://smcleod.net/2015/07/delayed-serial-stonith/
Comment linking /2015/07/delayed-serial-stonith/
on “Graceful Shutdown in Go: Practical Patterns”Submitted The HP 4951C Protocol Analyser as a story · 1 points
HN discussionImporting a Windows 10 shudder OVA _directory_ export from VMware Fusion worked perfectly and was quick to complete. - After installing the guest addons which was very quick as they must be quite light weight, all drivers worked perfectly and performance seems good in general (for windows anyway), have not tested 3D yet. - No native resolution scaling when resizing the VM window even after installing tools _NOTE:_ I…
Comment linking /
on “Native virtualization for macOS”I had a go at writing a bit of a sampling guide for Ollama/llama.cpp as well recently, open to any feedback / corrections - https://smcleod.net/2025/04/comprehensive-guide-to-llm-sampl...
Comment linking /2025/04/comprehensive-guide-to-llm-sampling-parameters/
on “Dummy's Guide to Modern LLM Sampling”Today I ran some perplexity benchmarks comparing F16 and Q8_0 for the K/V, I used Qwen 2.5 Coder 7b as I've heard people say things to the effect of Qwen being more sensitive to quantisation than some other models. Well, it turns out there's barely any increase in perplexity at all - an increase of just 0.0043. Added to the post: https://smcleod.net/2024/12/bringing-k/v-context-quantisatio...
Comment linking /2024/12/bringing-k/v-context-quantisation-to-ollama/
on “Bringing K/V context quantisation to Ollama”Just recently I wrote a brief post about how the title of ‘DevOps Engineer’ (and similar) were irking me, especially when looking for new roles / opportunities. DevOps isn’t a job title - it’s a culture and making click-bate / hipster job titles isn’t helping recruiters or business gain the value or plug the gaps they need: https://smcleod.net/tech/2019/08/08/camels-and-unicorns.html
Comment linking /tech/2019/08/08/camels-and-unicorns.html
on “What I Do as a DevOps Consultant”I'll have you some benchmarks next week / the week after or so. I'm building custom storage servers and will be writing it up as I go: http://smcleod.net/building-a-high-performance-ssd-san/
Comment linking /building-a-high-performance-ssd-san/
on “Endurance experiment kills six consumer SSDs over 18 months”I wouldn't suggest using hdparm for benchmarking, have a look at fio, it's easy to use and revised very detailed and accurate information - https://smcleod.net/tech/2016/04/29/benchmarking-io/
Comment linking /tech/2016/04/29/benchmarking-io/
on “Upgrade Raspberry Pi 4 with a NVMe boot drive”... And it was launched by Australia's Prime Minister I work @infoxchange as the operations lead, when I first heard the idea of a website or app for people that have found or are worried about finding themselves homeless in Australia I really didn't think it made sense - until I saw the stats showing that almost 80% of homeless people in Australia have regular access to a smart phone and data either via a cellular…
Comment linking /search-a-journey-of-delivery-on-a-budget
on “Show HN: AskIzzy – a mobile website for homeless people”NVMe is one of the most important changes to storage over the past decade. I'm currently replacing all our SANs with storage servers filled with NVMe SSDs (as the tier 1 storage, commodity SATA SSDs for second tier). I've posted the link to my first blog post about it in the comments on another NVMe post in the past: http://smcleod.net/building-a-high-performance-ssd-san/ I'm close to writing the next post around th…
Comment linking /building-a-high-performance-ssd-san/
on “NVMe, the fast future for SSDs”I'm currently working on a project to replace our older SANs with SSD-only storage servers - I've performed a few POCs with great results and am now documenting the build as I go. Not only is it going to give us some much needed IOP/s it's also going to save us hundreds of thousands of dollars on storage over the next 3 years. If you're interested take a look at: http://smcleod.net/building-a-high-performance-ssd-san
Comment linking /building-a-high-performance-ssd-san
on “1-2 year SSD wear on build boxes has been minimal”Actually we've found quite the opposite, because proprietary vendor management tools are generally hard to automate and integrate with existing frameworks such as puppet - the time it takes to administer them is far greater than any of our more standard systems. I just replied to a similar observation from someone else here: http://smcleod.net/building-a-high-performance-ssd-san/#comm...
Comment linking /building-a-high-performance-ssd-san/
on “Building a high performance SSD SAN, part 1”Absolutely agree, I migrated us from XenServer 7.2 to XCP-ng 7.6 without issue and the community has been great. https://smcleod.net/tech/2019/02/11/hello-xcp-ng.html
Comment linking /tech/2019/02/11/hello-xcp-ng.html
on “What Citrix is doing with Xenserver”I work for a 'tech' non-profit that I would certainly call successful. We have been around for 25 years and had many technological 'firsts' in Australia including I believe the first search engine and service directory. But have experienced substantial growth over the past 3-4 years and have in many ways become leaders in many areas both what we provide (products/services) and how we provide them (our internal platf…
Comment linking /talk-24-months/
on “Ask HN: Is it possible to start a successful tech non-profit?”I'd argue that a huge portion of the working population have for a long time exhibited an inability to think - probably in part due to the meaningless, widget churning work they're tasked with - I don't think that'll change, however being empowered to learn and try things beyond your individual grasp by leveraging AI as for example a coding parter can unlock whole new ways of thinking in domains they'd otherwise not…
Comment linking /2025/03/the-democratisation-paradox-what-history-teaches-us-about-ai/
on “AI's Biggest Threat: Young People Who Can't Think”Here's a HP 4951C I had on loan until recently: https://smcleod.net/tech/2017/12/27/hp-4951/
Comment linking /tech/2017/12/27/hp-4951/
on “Vintage HP computer systems”
