I reflected on why AI‑accelerated coding does not yield a 10× boost in shipping speed, grounding the argument in Amdahl’s law and highlighting the shift of bottlenecks to downstream stages.
When coding is 30 % of the end‑to‑end cycle and AI makes it 10× faster, the overall speedup is 1 / (0.7 + 0.03), which is about 1.37x.
If coding accounts for half of the cycle, AI‑driven acceleration yields roughly an overall speedup of 1.8x.
Faster code generation increases the load on every stage after it, so the review queue grows and the bottleneck shifts.
I fine‑tuned Qwen2.5‑3B for a production incident‑triage workflow :) GPU going brrrrrrr.
It encrypted everything so that only the attacker could decrypt it, then uploaded it to a /healthcheck URL, making the traffic look like a routine health check.
Keep each cleanup agent narrow. One removes dead code, one merges duplicates, and one updates stale docs and comments.
Run the same query on different engines, for example, PostgreSQL, SQLite, and DuckDB, and compare the result sets.
I decided to stop live‑streaming my fine‑tuning attempts after three tries because the content wasn’t dense enough for the audience.
Evals are tricky to get right, mostly because you can get them wrong in both directions. Too few, and you risk missing important failures. Too many, and you risk over-constraining the model. My mental model is simple - I ask one simple question: which failure mode does this eval catch?
px0 is optimized for exactly what we use IDEs for today: a quick skim and even quicker reviews :)
"For AI to be proactive, the database has to be reactive."
What we are missing is a common contract for events: how to publish them, subscribe to them, filter them, and resume consumption after a disconnection without missing changes.
On 2026‑10‑16 I posted a series of X replies that centered on three technical themes: (1) the importance of building the “right thing” – exemplified by Px0’s >1,000 daily sessions, (2) adopting XML guardrails plus explicit‑instruction‑as‑text handling to sanitize malicious HTTP responses, and (3) the discovery that the HTTP header X‑AI can be used for prompt injection, prompting a decision to strip such headers before feeding data to LLM agents.
Px0 daily usage exceeds 1,000 sessions per day.
XML guardrails + explicit‑instruction‑as‑text handling "work well" for sanitizing malicious response bodies.
Prompt injection can be delivered via an HTTP header named X‑AI.
Some models, like Opus and Sonnet, can ignore such headers by default, but many others do not.
Decision: strip or ignore any X‑AI‑style headers before feeding responses to LLM agents.
I think @rabi_guha's claim about releasing OpenUI 1.0 is super cool and much needed.
Irrespective of whether software engineering is coming to an end or not, I liken it to a roller‑coaster ride and stress the importance of staying optimistic about technology.
I congratulate @liftplaycode wholeheartedly.
so... it has come down to this 🤦 https://t.co/xbGx1R215P
OpenUI 1.0 release claim endorsement = "super cool and much needed".
I want to have 1:1s with my agents now.
Loops or an “agentic skill framework” have not been effective for me, so I am not pursuing them further.
I treat mono‑repos as a marketing problem, answering “yes” to the question of whether they solve a problem, which reflects my view of their positioning.
Just because Anthropic replaced Git with its own version control system does not mean you should too :).
The Daily Diff (TDD) has published 104 editions as of 2026‑10‑13 15:47 UTC.
Each edition contains roughly ~20 high‑quality stories, blogs, and research papers.
I gave permission to @YashSolanki_ to build a new edition on other topics: “Yes, go for it, build one :) ofc there will be people interested in diff domains.”
Philosophical note: “Doing the work and appearing to be working are two different things.”
The newsletter is curated by blending manual story selection with AI‑assisted trimming of irrelevant items, keeping the edition concise yet comprehensive.
The new open source world is you close the PRs. You ask the author to explain to you in plain english what they want. Then you add it yourself.
Everyone: please keep your variable names self-explanatory
Never say you are not monitoring or not available, it usually stops midway, just go to sleep.
It is 3 AM, and this is the instruction I am giving to Claude. new way of working
FIRE number needs a bump to account for this.
Daily diff is https://t.co/IyLtJNHLWy and has been used for the last 2 months.
The “new way of working” tweet was posted at 21:30 UTC.
Webpage said Hello World and nothing else; the issue is now fixed after being caught up with office work in the middle of the transition.
AI can give us every answer, but we are losing the ability to ask questions.
These posts capture my mindset on availability, experimental night‑time workflows, resource planning (FIRE number), and the philosophical tension around AI’s answer‑providing power.
I am waiting for UBI.
I am done, bro :)
Cat is out of the bag, I am off to Mumbai… for good.
In September, GitHub saw 7.38 billion commits 🤯
px0 --update
Waiting period for @23prithviraj’s request: last 3 weeks
Tenure in Bengaluru: 11+ years
These updates capture my personal status, a product release command, and the scale‑driving redesign principles GitHub announced, all shared on X on 2026‑10‑07.
In mid‑October 2026 I replied on X about px0, sharing personal impressions, feature ideas, product decisions and promotional activity.
Tried px0 briefly—really impressed by how simple it is to use.
I finally gave px0 shot and instantly realised I don't need Cursor anymore for code reviews.
One small feature request: I read code by following the curly braces. Click on an opening brace, which also highlights the end brace. Makes reading code easy.
One idea for code understanding: I currently ask Codex to read a repo and generate interactive HTML architecture diagrams showing components, relationships, and data flows.
I am contemplating. I have a working prototype but it is slightly heavy on browser, I am trying to optimize it before I make it live.
Browser is a better surface for IDE as compared to standalone app. This makes it easier to facilitate remote sessions.
Because I do not want multiple surfaces to monitor. Keeping things simple.
GitHub Issues - https://t.co/NvcnXc9xY0
that is precisely the thesis I am running with :) no one wants to look at the code, but still needs an assaurance that "nothing will break" --- enabling this is what px0 is optimizing for.
Enrollments are now open for Applied AI - Cohort 6, in case you are interested.
Cohort 6th starts on 14 November; details, curriculum, testimonials - https://t.co/JhtIHMrLbg
pro tip: do not let AI take away the serendipity in your life. make a concious effort to stumble upon something interesting every single day.
I noted on 2026-10-04 12:58 UTC that economics will always favour the machine.
This perspective frames my thinking about automation's systemic bias in future economic analysis.
In early October 2026 I posted a series of X replies about the prototype IDE px0, my three research approaches to code‑review automation, paper curation, and a stuck Progressive Web App.
Decision: Not really. I am focussing on post code generation.
Paper selection rule: I do judge a book by its cover and a paper by its title (and abstract) :)
PWA status on 2026‑10‑02: still at page with 18th. Tired refreshing and opening and close app. Still stuck on 18th.