André Terroir @terroir.systems · 01/10/2026Going even further - everyone should do it, regardless of the level of experience, if they wish to be able to build solid systems with AI assistance. 030
André Terroir @terroir.systems · 28/09/2026I'm also almost out of my OpenCode Go and SuperGrok subscriptions 🫠 001
André Terroir @terroir.systems · 28/09/2026This is especially true if you care about the quality of the code, maintainability, correctness, performance etc. 000
André Terroir @terroir.systems · 28/09/2026Besides, it's not fun to review bad code or steer an agents bumping into walls all the time. 000
André Terroir @terroir.systems · 28/09/2026After spending a weekend using Amp, I'm onboard - constraining token usage is slowing things down and not using frontiers models is limiting agent autonomy. Ultimately, it's a waste of invaluable human attention required to get to desired results. 300
Reposted by André TerroirDavid Crawshaw @crawshaw.io · 27/09/2026When in doubt, build something. 2222
André Terroir @terroir.systems · 27/09/2026It's interesting to see how the data centers of the cloud hyperscalers are distributed (or rather concentrated) across the world. gist.github.com/andreterroir...gist.github.comCloud topology snapshot, 2026-09-27Cloud topology snapshot, 2026-09-27. GitHub Gist: instantly share code, notes, and snippets. 000
André Terroir @terroir.systems · 26/09/2026“If you're not having fun, what's the point of this, really? […] If you're not enjoying yourself, go find somewhere where you're enjoying yourself.” - Adam Leventhal oxide-and-friends.transistor.fm/episodes/hir...oxide-and-friends.transistor.fmOxide and Friends | Hiring Processes with Gergely OroszBryan and Adam were joined by Gergely Orosz, the Pragmatic Engineer, to talk about Oxide's hiring process, the experiences that led to that process, and hiring generally. There's a lot there for an... 000
André Terroir @terroir.systems · 25/09/2026Actually, it’s kinda possible already with OpenCode’s service accounts. 000
André Terroir @terroir.systems · 25/09/2026It should be possible to give unused tokens to friends. 100
Reposted by André TerroirGergely Orosz @gergely.pragmaticengineer.com · 25/09/2026"Application development did not change much for 20 years, but now, a new wave is crashing in. Most computers in the future must be put to work at least in part w/o programmers. Not surprisingly, programmers will instinctively resist methods described in this book." - 1982, James Martin (cont'd) 2747
André Terroir @terroir.systems · 25/09/2026This might no longer be the case, since now it’s possible to bring your own model access. I’m looking forward to give it a try with Open Code Go. ampcode.com/news/free-ag...ampcode.comFree AgentAmp is now free to use when you bring your own compute and model subscriptions/keys. 000
André Terroir @terroir.systems · 25/09/2026Coding agents are like Linux distributions - there’re a lot of them, many are amazing and none are perfect. 200
André Terroir @terroir.systems · 23/09/2026This is very cool tracking of global networking. geocables.comgeocables.comGeoCables - Live Submarine Cable Encyclopedia (716 cables, real-time latency)Interactive map of 716 submarine cables worldwide. Live routes, real-time latency from our measurement network, and operational status. 000
André Terroir @terroir.systems · 22/09/2026Models having an explicit notion of planning is a natural progression. I’m looking forward to seeing what comes out of Advanced Machine Intelligence Labs. amilabs.xyzamilabs.xyzAMI Labs: Real World. Real Intelligence.AMI - Advanced Machine Intelligence - builds world-model-based AI that understands the real world. We develop safe, controllable intelligent systems for industry, robotics, healthcare, and beyond. 000
André Terroir @terroir.systems · 22/09/2026Decision models is a step in the right direction. It’s an example of how to achieve significantly better efficiency and, likely in the future, quality by better fitting a model to the problem. Another plus - we drop anthropomorphic qualities of the output. simonwillison.net/2026/Sep/21/...simonwillison.netJev introduces a new shape of LLM—System One, aka Decision ModelsLast week TypeSafe AI unveiled Jev, their first example of a new category of model that they are calling “System One models” (I’m with Maggie Appleton, I think “decision models” … 100
André Terroir @terroir.systems · 28/08/2026In the time I've been contemplating these ideas, LinkedIn announced Northguard and Apache Pulsar built scalable topics. I hope my research is still useful if you're interested in distributed message logs and stream processing. terroir.systems/can-kafka-su...terroir.systemsCan Kafka Support Elastic Partitioning?The Promise of Dynamic Partitioning Partitioning is a recurring challenge around building systems on top of Apache Kafka distributed message log. Some com... 000
André Terroir @terroir.systems · 20/08/2026A good follow up read: beej.us/blog/data/ai...beej.usOn Making 001
André Terroir @terroir.systems · 20/08/2026It's not that there's nothing to learn, but I'd rather spend my limited energy on the best ideas of people who invested their valuable time into developing them, rather than digging through the statistical averages of AI-produced sameness. 000
André Terroir @terroir.systems · 20/08/2026That's probably why many AI-written projects are uninteresting beyond their high level ideas. If the author didn't care enough to put in the effort into building the system, why would I spend my limited time exploring it? 100
André Terroir @terroir.systems · 20/08/2026As an individual engineer, your value is not in the ability to type the code but in the thinking behind and in the good judgement beyond the statistical averages of an LLM. And ability to write the code is the foundation for all the surrounding software expertise: haskellforall.com/2026/05/type...haskellforall.comType out the codeFreecoding improves broader programming proficiency 010
André Terroir @terroir.systems · 20/08/2026On the other hand, the value provided by a software business extends far beyond the effort required to write the code, which is never the bottleneck for a high-leverage product (@joran.tigerbeetle.com): shows.acast.com/software-uns...shows.acast.comTigerBeetle's Spectacular Jepsen Report - with Joran Greef | Software Unscripted 100
André Terroir @terroir.systems · 20/08/2026Is AI-created software valuable? How does the cost of writing code influence the value it provides? On one hand, the effort put into making the software valuable makes it more prized by the creator and, by extension, more valuable to its users (@jola.dev): jola.dev/posts/no-cos...jola.devNo cost, no valueWhy generating code makes it meaningless, why writing code by hand has value, and why toil is a critical part of the human experience. 200
André Terroir @terroir.systems · 20/08/2026Some thoughtful deliberations on AI and ownership of work: jola.dev/posts/a-comp...jola.devA computer can never be held accountableFor better or for worse the things you create with LLMs are your responsibility, just like with any other tool. 100
André Terroir @terroir.systems · 20/08/2026Just set-up external monitoring for my websites with larm.dev. It's been a delightful experience. Here's my external dashboard: terroir.status.larm.devlarm.devLarm | Uptime Monitoring for Engineering TeamsKnow before your customers do. Uptime monitoring with full request traces, multi-location verification, and status pages. Free to start. 021
André Terroir @terroir.systems · 17/08/2026It would be interesting to hear your thoughts on the Qwen-derived Bonsai models, as they address the performance aspect specifically. A ternary Bonsai 27B feels both impressively capable and performant enough (on a 32GB M2 Max). prismml.com/news/bonsai-...prismml.comPrismML — Announcing Bonsai 27B: The First 27B-Class Model to Run on a PhoneToday we're announcing Bonsai 27B, our multimodal flagship: ternary at 5.9GB for laptops, 1-bit at 3.9GB for an iPhone 17 Pro, with a 262K-token context. 020
André Terroir @terroir.systems · 15/08/2026The commoditization of access to both offensive & defensive cybersecurity capabilities finally creates an incentive for secure software by lowering the cost of exploiting & mitigating vulnerabilities simultaneously. The risks and likelihood of breaches increase, while prevention efforts decrease. 000
André Terroir @terroir.systems · 15/08/2026Right now cybersecurity is a race to discover and exploit/mitigate vulnerabilities. It doesn't have to be if software is secure by construction and exhaustively tested before being released. 000
André Terroir @terroir.systems · 15/08/2026It's good to hear someone having a sober conversation about the threats of AI, including bioweapons. What a coincidence that the fearmongering is so well aligned with the business incentive of the big AI labs. oxide-and-friends.transistor.fm/episodes/the...oxide-and-friends.transistor.fmOxide and Friends | The Open Weight Revolution with Simon WillisonSimon Willison joined Bryan and Adam to discuss a wild couple of weeks in AI. First, we have a high-profile, AI-induced, AI-diagnosed security incident. Next we have the rise of frontier-class open... 010
André Terroir @terroir.systems · 15/08/2026Contrary to the narrative of the big AI labs, open-weights models and democratizing access to AI is essential in the transformation to the transition to secure software. Continued gatekeeping of cybersecurity capable models would hold the world hostage. 110
André Terroir @terroir.systems · 15/08/2026We’re in for some turbulent times for sure, but ultimately we will come back better off on the other side. 000
André Terroir @terroir.systems · 15/08/2026The recent AI-related cybersecurity incidents make me hopeful about the future. Secure software without vulnerabilities is possible. The bottleneck was always identifying security flaws, which contemporary models can remove. www.foreignaffairs.com/podcasts/cyb...foreignaffairs.comCyberwarfare in the AI AgeA Conversation With Jen Easterly 310
André Terroir @terroir.systems · 14/08/2026Bonsai 27B is the first local LLM that's actually useful with a coding agent on my 32GB M2 Max MacBook. The models I've tried earlier were either too prone to mistakes, too slow, or too resource hungry. prismml.com/news/bonsai-...prismml.comPrismML — Announcing Bonsai 27B: The First 27B-Class Model to Run on a PhoneToday we're announcing Bonsai 27B, our multimodal flagship: ternary at 5.9GB for laptops, 1-bit at 3.9GB for an iPhone 17 Pro, with a 262K-token context. 020
André Terroir @terroir.systems · 11/08/2026The complexity of having to manage Pulsar itself (brokers), a log store (BookKeeper), a metadata store (Oxia/ZooKeeper), and Pulsar proxies might be a factor in lower adoption. Kafka's ubiquity is likely another. 010
André Terroir @terroir.systems · 11/08/2026How come you don't hear more about Apache Pulsar? Geo-replication, multi-tenancy, tiered storage, and especially scalable topics (preview release) sounds very interesting. pulsar.apache.org/docs/5.0.x/c...pulsar.apache.orgScalable topics | Apache PulsarUnderstand scalable topics (Topics v5), the topic type that grows and shrinks at runtime by splitting and merging key-range segments. 100
André Terroir @terroir.systems · 28/07/2026The idea behind semantic caching is to leverage a cache when the common prompt prefix is virtually the same, but not exactly byte-to-byte identical. docs.nvidia.com/deeplearning...docs.nvidia.comSemantic Caching — NVIDIA Triton Inference Server 000
André Terroir @terroir.systems · 28/07/2026The cache is typically stored in GPU memory, which is expensive, but it's also possible to offload it to the main memory. nvidia.github.io/TensorRT-LLM...nvidia.github.ioKV cache reuse — TensorRT-LLM 000
André Terroir @terroir.systems · 28/07/2026Because prompt caching only works with exact prefixes of the message history, it's most effective to include the static data, such as tool descriptions, system prompt and standard instructions at the beginning of the context. 100
André Terroir @terroir.systems · 28/07/2026Without caching, to produce a next response the model has to process the whole message history first . Resuming from a cache allows to restore the model state at the point before a next response can be produced, without having to do the expensive processing work. 000
André Terroir @terroir.systems · 28/07/2026LLM prompt caching is unlike data caching, were the result of an operation is served from the cache to avoid a more expensive computation or lookup. Instead it's the internal state of the model (a bunch of multidimensional vectors) that's cached, allowing to resume inference from a previous point. 300
André Terroir @terroir.systems · 27/07/2026Model reasoning has costs: in output tokens, context window space, increased response time and time to the first response token. 000
André Terroir @terroir.systems · 27/07/2026Some proprietary models support continuity of reasoning, in which case the actual reasoning chain returned to the client is encrypted, to keep the client stateless. For the new Anthropic models thinking blocks can be redacted. platform.claude.com/docs/en/buil...platform.claude.com 000
André Terroir @terroir.systems · 27/07/2026Depending on the model, the reasoning blocks could also be interleaved with responses, including tool calls. Proprietary models from the big labs do not return actual thinking tokens, but summaries. developers.openai.com/api/docs/gui...developers.openai.comReasoning models | OpenAI APILearn how to use OpenAI reasoning models in the Responses API, choose a reasoning effort, manage reasoning tokens, and keep reasoning state across turns. 100
André Terroir @terroir.systems · 27/07/2026Typically thinking precedes a response and is not included into the conversation history sent to the model on each turn. The goal is to improve quality of each individual response. As reasoning increases the response length, including it would exhaust the context window very quickly. 110
André Terroir @terroir.systems · 27/07/2026Reasoning (also thinking) LLMs are the models that have been trained (or fine-tuned to be precise) to output "chain-of-thought" tokens before producing a response, increasing its quality for certain classes of problems. www.ibm.com/think/topics...ibm.comWhat Is a Reasoning Model? | IBMA reasoning model is a large language model (LLM) fine-tuned to break complex problems into smaller chain-of-thought (CoT) steps, often called “reasoning traces,” prior to generating a final output. 200
André Terroir @terroir.systems · 24/07/2026An RFD/ADR process sounds especially useful in a fast-paced environment when you don’t have time for in-depth exploration via design docs but decisions piled up quickly. 000
André Terroir @terroir.systems · 23/07/2026Higher productivity != more code. Higher productivity is higher level abstractions. 000
André Terroir @terroir.systems · 23/07/2026Until we figure out how to teach agents software design, building well abstracted components, that can be understood independently and combined safely, sounds promising. The alternative is oftentimes a tightly coupled snowball of code. maheshba.bitbucket.io/blog/2026/07...maheshba.bitbucket.ioThe Bottlenecks for AI-Driven System Design: why Principal Engineers are not (yet) obsolete.I’ve been building systems, teaching system design, and writing papers about novel designs for a couple of decades now. In my view, agents are magical self-writing distributed programs: my personal re... 100
André Terroir @terroir.systems · 23/07/2026In my experience, the time invested into defining the interfaces and module boundaries is very much worth it. It's a good way to ground the model-generated code and get faster to the target result. 000
