developer relations at @aiand_ | agentic AI enthusiast, building in public. | prev: @GoogleDeepmind, Google Stadia
Tokyo, Japan
Joined May 2008
- Tweets41.8K
- Following6.9K
- Followers2.5K
- Likes224K
Jetha Chan retweeted
I'll be presenting two posters at AACL2026!
WAON: The largest Japanese image-text pair dataset
📅 Nov 6, 11:00–12:30
HakushoBench: A challenging Japanese chart and table VQA benchmark
📅 Nov 7, 13:30–15:00
If you're interested in benchmarks & datasets, I'd love to chat!
WAON, the largest Japanese image-text dataset, and HakushoBench, a challenging Japanese chart and table VQA benchmark, have been accepted to #AACL-IJCNLP2026 Findings! 🎉
We hope you’ll find them useful for training and evaluating VLMs.
See you in Hengqin! 🇨🇳✈️
Jetha Chan retweeted
One of my projects has the problem that the sub is cheap, and it's easy for 1-2 heavy users to blow the profit from multiple subs by incurring too much CPU hours. I don't want to raise prices and I don't want to set limits (yet)
So I asked Gemini 4 Argon to do deep review of the infra and evaluate a few ideas against actual production numbers to cost cut. It discovered a way to reduce spending by 91% / 17x less by switching parts of my architecture and targeting a different provider!!
It's a pretty huge refactor on a production system, but 17x less is making it worth it to sink a day into this.
Highly recommend doing the same with your projects now and then
Jetha Chan retweeted
DGX Spark setup and config: we spent a day on it and wrote a guide for Spark owners and their agents. Kernel, firmware, memory, power cap, two open kernel bugs and how we patch around them. Thanks @FT_IOxCS for the CVE heads-up.
al-engr.com/dgx-spark-how-we…
Jetha Chan retweeted
I have autonomous agents watching the logstream and events of all of my services to identify weird traffic or problems, and proactively file issues and kick off other agents to push a fix
I woke up to a PR adding a setup guide for @MistralAI, because the watcher agent saw the `mcp_client_authorized` event that get's emitted when a user connects Shard Connector to an AI client, and decided to hand-off an implementation
This is honestly really cool, and the fact that I don't even have anything in my instructions to tell it to do that makes this even more surprising
Jetha Chan retweeted
RIP Popular Mechanics
The lead story in the new issue of Popular Mechanics is “How to Fight a Data Center.”
How sad and bizarre. Another once-great technology magazine turned against technology by leftists.
Subscribe to @arenamag - we actually like technology, and America 🇺🇸
Jetha Chan retweeted
Spent the whole night in Discord helping people with the NIKKE launcher malware 💀
If you don’t know where to start, at least protect yourself: stop opening the PC launcher, change your passwords from your phone, log out of all devices.
gist.github.com/Diew/ca29440…
#NIKKE
Jetha Chan retweeted
In a weird way, everybody is rooting for you. When you go to a job interview, the interviewer wants you to be amazing so they don’t have to look anymore. when you go on a date, the person wants to have an amazing date with you. If you make a piece of content, people want to see amazing pieces of content and they have no problem if it’s yours.
the instant you get into the mindset that people want you to do well, your entire world gets better. And you’ll be shocked to know this pattern keeps repeating in different areas of life.
Jetha Chan retweeted
Cities, cars, trucks, busses, trains, headlights, break lights, turn signals, water, clouds, rain, bridges, < 500k, no libraries #webgpu
greggman.github.io/gen-city-…
Jetha Chan retweeted
Dude had a NordicTrack treadmill ($2-3k) and wanted to use massive screen to watch Netflix.
NordiTrack locked that option behind $39 a month membership for its iFit app.
Claude figured out screen was Android 9 tablet and “found that iFit’s motor service runs a small server on the console that apps can talk to”.
It disabled iFit’s apps one at a time and the “real fix was removing iFit’s lockdown and updater app for my user (reversible). It turned remote access off after every restart, re-enabled everything I’d disabled, and ate a big share of the CPU. Without it the treadmill boots unlocked and the tablet can finally breathe.”
He can now CrunchyRoll and Netflix without paying an extra $39 a month on top of those streaming subs lolol
Jetha Chan retweeted
life update:
Jetha Chan retweeted
Modding built a culture around taking software apart and changing what its creators intended. Now we’re policing which tools modders can use to do it? That’s a funny place to start moralizing.
Nexus Mods will make AI disclosure mandatory for all mods, but won’t ban AI-generated content.
The platform announced several changes following community feedback:
> Authors must declare AI usage and explain how it was used in AI-assisted mods.
>New rules will let authors control AI use, AI training and derivative mods. Permissions will default to “Ask me”.
>Content filters: Users will get easier options to hide AI-generated mods.
> AI-tagged mods received just 2.13% of all Donation Points this year. Reward rules won’t change for now.
Nexus will also never use mods to train AI or license them for that purpose.
Existing mods have until January 2, 2027, to meet the new requirements. After that, violations could lead to account restrictions or mod removal.
Jetha Chan retweeted
Just a reminder that DeepSeek V4.1 Flash 1017, GLM 5.3 Flash, Qwen 3.8 Next Flash, and even Qwen 3.8 27B are all outperforming (in both intelligence and capabilities) every model that was considered "frontier intelligence" in Xmas 2025 (just 11 months ago)
Opensource AI is on fire
Jetha Chan retweeted
Watched the latest @LinusTech video and they are running QWEN 3.8 27B with pipeline parallelism GGUF on LM studio Q4KM and complaining it isn’t fast.
Oh my eyes.
Please PLEASE someone who’s close to
him educate him on VLLM and Strata.
He could literally be getting 5x better speeds and better intelligence.
Sure the setup is complex but it’s a one time setup with a cloud agent.
Jetha Chan retweeted
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list.
Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%.
Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything.
Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO.
h/t @OliverMolander
Would you move your team off frontier APIs if the bill dropped 90%?
Jetha Chan retweeted
80,000 lines of code. 92,000 trees. Real-time ray tracing.
Two weeks of development using Claude Code to build a complete off-road racing game from scratch on a custom C++/Vulkan engine.
Source:
github.com/markaren/threepp
Jetha Chan retweeted
‼️ BREAKING: Google says Android TV box maker Videostrong infected its own users and sold the access to criminals.
They built a backdoor into the firmware of 1.3M+ set-top boxes. If a user deleted it, it came back on the next reboot.
Buyers turned the boxes into residential proxies and ad-fraud bots.
Google has banned Videostrong from certified Android and says the takedown also cuts off at least 10M more infected devices.
