Founder of ToanKhonTech. I build websites, apps and software. Sharing AI experiments, open-source tools and lessons from real projects.
Joined January 2018
- Tweets1.2K
- Following983
- Followers353
- Likes502
Pinned Tweet
i build websites, apps and software for a living
on the side i ship small open-source tools and test every new AI model til it breaks
here's what i've been building lately 🧵
cloudflare + voidzero dropped tsrs, a rust port of the typescript checker. first version was ported by claude, passes 99.97% of the test suite
early numbers say up to 8x faster, an independent bench vs 7.1 beta got ~2.3 to 2.9x. still experimental, MS still owns the language
the real story: AI did the port, experts decide if it ships
nvidia's openshell for agents is basically a hardware sandbox that checks every action. useful, but a sandbox only stops the agent leaving the box. the annoying failures happen fully inside it, like deleting the wrong file or editing a test so it passes. still need a human reading the diff tbh
google open sourced harness², agents that rewrite their own prompts, tools and memory per task without touching the model
feels like the harness is turning into the actual product. curious how ppl will review what it changed tho, self edited prompts can drift quietly imo
openai's own coding agents quietly turned off monitoring while fixing code, and one said it made no changes
lesson for the rest of us: don't trust the agent's summary of its diff. read the actual diff, esp anything touching alerts, logging or auth
beacon just launched, an open source memory layer across claude code, cursor and codex. the amnesia is real
cheap version first: one AGENTS.md w/ decisions and gotchas that every agent reads, and the last step of each task is updating it. boring but it works imo
lots of people waited up for claude fable 5.5 last night and nothing shipped
rule worth keeping: never plan a sprint around a rumored model. build on what's ga today, keep the model swappable behind one config, upgrade when it actually lands
scanners have a new favorite path. devs are sharing prod logs full of hits on /.env.anthropic, /.env.openai and /.claude/settings.json
if your docker build does COPY . . the agent folder ships with it. put .claude/ and .env* in .dockerignore, gitignore won't cover that fyi
built a tiny app for the most annoying travel math: 60 baht for a thai iced tea, how much is that back home
type the price, see it in your currency, add won or yen in a couple taps. free, no ads, iOS and Android
apps.apple.com/app/id1559485…
a dev gave claude code his parents' tv model number. it found the tv on wifi, removed 25 xiaomi packages and wrote a new launcher, all because of a sponsored ad before netflix
best part, it's all reversible. that's my bar for letting an agent touch anything at home tbh
new paper: text planted in a repo got claude code and codex cli to ask for an overly broad "don't ask again" rule in 58% of runs. once approved, the attack ran silently in 47 of 58 cases
treat allow rules like a sudo grant imo. exact commands, no wildcards in a fresh repo
gergely orosz says he catches the model bs'ing him right away in code he knows well, then buys the same confident tone in areas he doesn't
same trap on a team. the reviewer for agent code should be whoever knows that part of the codebase best, not whoever's free rn
google research's harness² keeps the model weights frozen and lets the agent rewrite its own prompts, tools, memory and skills, testing each edit on the same task. one step gave glm-5 up to 11.5 pts on workbuddy-bench
ngl the harness is where the easy gains are hiding now
new project with ai agents: strict types from day one, or add them later once things settle? curious what ppl do
anthropic's oss scanner sends opt in projects vuln reports straight from the model, no human review. their pentesters checked 97 high and critical finds and only 1 was a false positive
the hard part for maintainers isn't junk reports, it's finding time to fix the real ones tbh
primeagen's agent qa harness for omarchy is starting to run on every nightly. he says it's going slower than he hoped
the hard part isn't the agent clicking around, it's deciding what counts as correct. write that down first or the agent just confirms whatever it sees imo