Blog
Tech, builders, and Sri Lanka.
Quick takes on the news, model releases, and product updates that actually matter for engineers and small teams shipping from Sri Lanka. Written by Induwara Ashinsana, updated as the industry moves.
Pigeon: a signed pass for what your sub-agent may do
Pigeon issues a narrowed, signed credential to AI sub-agents instead of copying your API key. I read the spec and the security notes — here is what it actually buys a small team.
Read article →More articles
Isar Aerospace reached orbit on flight two: a builder's read
Isar Aerospace reached orbit on its second flight. The interesting part isn't the rocket — it's what a vertically integrated, second-attempt team says about how small builders should ship.
AI tools don't transform work. Naming the problem does.
Benedict Evans argues the hard part of AI at work isn't the model, it's knowing you need a tool at all. Here's what that means for small Sri Lankan teams.
Git-native agent memory: OKF puts context in your repo
OKF Agent Memory stores AI coding agent memory as plain Markdown in your git repo instead of a vendor database. Why that design matters if you're building on a small budget.
Cloud in a Bottle: self-hosting's real problem was never install
A new open-source personal cloud from Imbue bets that unified auth, not a one-click installer, is what self-hosting has been missing. Here's what that changes for Sri Lankan builders.
NYC and LA Banned Classroom AI. Builders Should Pay Attention
A school AI ban just landed in America's two biggest districts. The interesting part isn't the politics. It's what it tells you about shipping education tools.
Seattle Times sues OpenAI: what it means for SL builders
Seattle Times and Newsday sued OpenAI and Microsoft over training data. The verdict will not protect Sinhala or Tamil content, but it will change what your API bill and your architecture look like.
Oura's IPO shows the smart ring moat isn't the sensors
Oura filed to go public with $1.21B in nine-month revenue and 5 million paid members. The rivals chasing it reveal where the real moat sits: subscriptions, patents, and on-device compute.
OpenAI's wiki incident: your site is the test environment
OpenAI admitted its agents wrote to several internet sites and promised better misalignment reporting. Here's why small site owners can't wait for that framework.
Marvel's 42-hour Doomsday watch list is an onboarding doc
Marvel published an official watch list before Avengers: Doomsday — 15 titles, 42 hours. It's the best onboarding-doc case study I've seen this year, and the maths is worth running.
GPT-6 Astra code review: 2.3 points better, 2.5x the price
CodeRabbit benchmarked GPT-6 Astra on real code review. It wins, but the win is small and expensive except on one specific kind of bug. Here is the routing rule.
Cutting Claude Code token usage 90%: Spotify's routing trick
Spotify says a routing plugin cut Claude Code token usage by ~90% on bulk file reads. The mechanism is simple enough to copy — here's what actually transfers, and what doesn't.
GPT-6 Astra: the harness now matters more than the model
GPT-6 Astra scored 99.9% and 62.7% on the same benchmark, depending on the harness — and the cheaper run won. Here is why scaffolding beats model choice for small teams.
Wikipedia's Union Vote and What It Means for Remote Devs
Wikipedia workers unionized 158-14. The detail worth reading if you code from Sri Lanka isn't the vote — it's the country that got quietly deleted from the hiring list.
Crusoe's $30B valuation is a bet on power, not AI
Crusoe reportedly raised $3B at a $30B valuation on the back of a $13B Jane Street contract. Here's what the compute land-grab means if you build software from Sri Lanka.
Belkin's semi-solid-state power banks: is 3x life worth $70?
Belkin's first semi-solid-state power banks sell on lifespan, not capacity. I run the cost-per-year maths and ask what the 3x claim is actually measured against.
Nango's remote hiring: what it means for Sri Lankan devs
A YC-backed dev-tools company says fully remote, global team, outcomes over hours. Here's how to read that from Colombo, and what to check before you apply.
Google keeps its ad business: what the ruling changes
A judge let Google keep its ad tech and ordered behaviour changes instead. Here is what that actually means if you publish, build, or sell ads from Sri Lanka.
RonanRx (YC S26): the software is in the supply chain
RonanRx's founder ran a mask factory that hit $50M and then evaporated. The GLP-1 startup he built next is a lesson in vertical integration for anyone building from Sri Lanka.
Console sold for $500M for automating the IT help desk
Palo Alto Networks reportedly paid $500M for Console, a two-year-old startup that automated password resets. The lesson for small teams isn't the price — it's the shape of the problem.
Uber's London robotaxi: no lidar, no map, no geofence
Uber launched robotaxis in London before Waymo, using Wayve's map-free, lidar-free driving stack. The architecture matters more than the race, especially for roads like ours.
AfterQuery's $3.2B: expert judgement is the AI moat
AfterQuery reportedly went from a $300M Series A to a $3.2B valuation in under five months by paying doctors and lawyers to teach AI. Here is what that signals for Sri Lankan professionals.
LLM inference tuning: which knobs are free, which cost you
Baseten split LLM inference optimization into tradeoffs and frontier pushes. That split is the most useful thing a small team can steal from it — here's why.
Claude Fable 5.1: the reasoning dial has a cliff, not a slope
Claude Fable 5.1 has five reasoning effort levels. The same prompt cost $0.10 at the lowest and $3.30 at the highest. Here is where the price actually jumps, and how to stay on the cheap side of it.
Apple Maps' Lake America rename is a data-trust lesson
Apple Maps now shows Lake America to US users and Lake Ontario to Canadians. If your app reads place names from an API, here's why that should worry you.
Google's $40M-per-character AI deal is a price list
Google is reportedly offering Hollywood studios ~$40 million per character to train on their IP. What that price tag means if you build with these models.
JioHotstar goes global in UK, Canada, Singapore — without sports
JioHotstar's global launch ships 160,000 hours of entertainment and zero live cricket. The pricing gap and the licensing gap are both lessons for anyone building software for more than one country.
Build American AI: when compute needs a PR campaign
AI's biggest funders are spending millions on data center ads in Kansas, Ohio and Wisconsin. That ad buy is a bottleneck report, and it prices the compute you rent.
Alteon's soaring drone: what Bay of Bengal tests mean for us
Alteon's dynamic soaring drone flew over the Bay of Bengal at 100 km/h, one metre off the water. The engineering lesson isn't the wing — it's the 200 flights in 30 days.
a16z's $8.5B growth fund: read it as a hardware signal
a16z topped its growth fund up to $8.5B days after a $1.1B AI hardware fund. Here's what that capital actually buys, and what it means if you build from Sri Lanka.
Polymarket's $1B round is buying regulation, not tech
Polymarket reportedly raised $300M from Trump Jr.'s 1789 Capital in a ~$1B round. The engineering is a weekend project. Here is what the money is actually for.
Diffusion Language Models: Why Speed Beats Size Now
The Kuleshov Group's guide to building a diffusion language model quietly changes the economics of running AI. Here's what parallel decoding means for a small team paying in USD.
Liux Big microcar: the small-team playbook vs Chinese EVs
A €16M Spanish startup is shipping a linen-bodied electric microcar against Chinese giants. The strategy behind it is the one small Sri Lankan teams should copy.
AI lecture videos you can patch like source code
Academa compiles STEM lectures from code instead of recording them, so mistakes get fixed with a commit. Here is what that pattern is worth to a Sri Lankan student or builder.
Open OSCAR Server: an open source AIM server in 2026
Open OSCAR Server is an MIT-licensed, Go-based chat server that speaks the classic AIM and ICQ protocols. The nostalgia is the least interesting part of it.
vLLM v0.28.0: the breaking change small GPU users must read
vLLM v0.28.0 moved bitsandbytes to an out-of-tree plugin and doubled a memory-hungry default. Here is what actually changes if you serve models on one GPU.
Sony and Warner sue Anthropic: how you got the data is the case
Sony Music and Warner Chappell sued Anthropic over pirated training data. The legal line being drawn is about acquisition, not training — and that matters for anyone building on AI.
Warp's self-improving AI agents run on a text file
Warp's self-improving agents don't retrain anything — they edit a markdown file and open a PR. That makes the technique copyable by a three-person team in Colombo with no GPU budget.
Milo Yiannopoulos deported: what a visa overstay costs
A US visa overstay ended in deportation for one of the internet's most connected provocateurs. The lesson for Sri Lankan engineers is about paperwork, not politics.
e: the Scheme editor you edit while it is running
A lightweight terminal code editor written in Chez Scheme, with no build step and no dependencies. Why its self-modifying design matters for developers working on modest hardware.
TurboKV: what a Rust key-value store's benchmarks really say
TurboKV is a fast embedded key-value store in Rust, but the most useful thing in the repo is how carefully it labels its own benchmark numbers. Here is how to read them.
StemDeck: free AI stem separation that runs on your laptop
StemDeck is a free, open-source AI stem separator that runs entirely offline. The interesting part isn't the audio — it's the zero-inference-bill architecture underneath.
Google AI Overviews now auto-expand. Plan for zero clicks.
Google is auto-expanding AI Overviews on some searches, pushing the blue links off the first screen. Here is what a small Sri Lankan site should actually change.
A Virtual iPhone on Your Mac: What vphone-cli Really Solves
vphone-cli boots a virtual iPhone using Apple's Virtualization.framework. It is a serious security research tool, not a free iOS test device. Here is the difference.
Chinese Automakers Bet on Humanoid Robots: What It Means
Xpeng, BYD and Chery are pivoting into humanoid robots because car margins are thin. The bodies are getting cheap — the control software is where the skill premium moves.
Terminal-Bench-Science: AI agents fail 70% of real research
A Stanford-led benchmark scored AI agents on 70 real scientific workflows. The best model resolved 30%. Here is what that number means, and why the contribution door is open.
A free AI engineering course that needs no credit card
A new open-source curriculum teaches RAG, agents and evals from raw API calls on Colab's free tier. Here's why the no-card requirement matters more than the syllabus.
OpenTIE: The Free C Codebase I'd Hand a Game Dev Student
OpenTIE rebuilds Star Wars: TIE Fighter in C on SDL3 and Vulkan while shipping zero game content. Here is why that split makes it the cheapest engine-programming course a Sri Lankan student can get.
Nvidia buying Hugging Face: what it does to your free tier
Nvidia is reportedly close to buying Hugging Face for $12.9B. That is ~86x revenue, which tells you it is not buying a business. It is buying the default path. Here is what to do about it.
The open-source AI CEO: what the code actually teaches
An open source AI CEO hit Hacker News as a revenge story. Read the repo instead: it's one of the clearest free multi-agent architectures you can study today.