Data - LLM - AI - Autonomous Agents - Blockchain
Bangalore - Sydney
Joined October 2007
- Tweets7.8K
- Following4.1K
- Followers669
- Likes2.9K
Durga Sahoo retweeted
There are seven main steps for your company to go from tool adoption to agentic org.
Most companies sit on step one, shopping for step seven technology. They need step four discipline.
Step one: tool adoption. AI tools exist across the org, humans run each one. Developer closes the lid, automations die. Somebody goes on vacation, the workflow goes with them.
Step two: context layer. Knowledge graphs per department, each behind its own API gateway. Sales queries engineering through an API. Separated domains keep agents from hallucinating across departmental boundaries.
Step three: harness engineering. Build evaluation criteria, routing logic, and quality gates before you pick a model. The LLM comes in last. Your harness is your IP and your capex. The model underneath is a commodity you swap when something better ships. If changing the LLM breaks your system, you built a dependency, not a platform.
Step four: go headless. Host every automation outside a laptop. It fires on schedule, on conditions, without a human logging in at 9 AM. Companies stall here because nobody forces the migration from “script on my desktop” to “service that runs on its own.”
Step five: governance. RBAC on every knowledge structure from day one. TTL on every agent. Dead agents are tech debt with active API keys. Past 100 tools, you need lifecycle management. Retire anything that hasn’t run in 90 days. Deprecate anything whose eval scores drop below threshold.
Step six: agent-to-agent. Stitch headless instances together within a value chain. Agent A finishes, hands off to Agent B. An LLM sits between them as a quality gate, judging every handoff before it passes through.
Step seven: step functions. Orchestrated, semi or fully automated workflows. Agents handle volume. Humans handle exceptions. The org runs on agent output.
We walked a VP of Engineering through all seven.
Told him: “You’re on step one. But over a hundred tools means step four is closer than you think.”
Durga Sahoo retweeted
It's 9:57 am. Been up all night. Postiz is back up!
Been a very hard night, and I feel like I let everybody down.
All the comments I got were very encouraging; nobody scolded me. Thank you so much for that!
During the day, a major PHP vulnerability was discovered, affecting about half the web.
Deployments on @railway stopped, and I assumed it was because of that, even though you do want deployments to upload the latest WordPress fix.
Everybody pinged Railway on Discord, and there was no answer- crickets... and people started getting extremely worried, especially because of >> PHP.
Then I got an email from Anthropic and Stripe saying my API keys were compromised. Still no answer from Railway, but I didn't take the risk; I rotated them and put Postiz down.
We all, even in Discord, already understood there was a leak, but Railway reported it only one hour later.
So the goal was to rotate all 140 keys (that's wild), because each key is duplicated across different containers (scattered). Some have the same variable with a different value (like different DB pool connections)
I was overwhelmed; it was too much, and I froze. I didn't know what to do.
I am lying on my Airbnb bed, but I know I need to get into sitting mode to be more serious, so I go to the kitchen and sit on a chair.
I asked claude to vibe code a system that:
- Will take all the keys and values from Railway (using their MCP)
- Will have an easy way to update all of them or one by one. Based on containers.
- Give me categories so I can be more organized. and prioritize.
I also started rotating keys from 12 am to 7 am.
All while Railway is still down and deployments are paused.
Once I finished I used Railway MCP to update all the env variables. if AI didn't exists I don't know how I could have done it.
Then, around 8:40am to 9:00am, we started deploying everything back to Railway to minimize downtime as much as possible.
And now everything is up.
I feel bad because DevRels at Railway sweet-talked me into staying after each incident and promised things wouldn't happen again; after two incidents, I should have migrated.
I couldn't find a platform I enjoyed as much as Railway, so I didn't want to switch.
Railway problems will happen again, and I will do everything I can to make sure they don't affect me like that anymore. The main problem was the lack of communication and 17 hours of downtime, which isn't realistic.
In the next few days, we will come up with a solid plan to migrate from Railway to a different service, trust is gone.
Also, Postiz grew from $300k to $303k MRR overnight. If you tried to cancel your subscription and couldn't, please contact [email protected], and they'll issue a full refund (super sorry about that).
This week, will be security week for everybody!
We are currently shutting down Postiz for all customers. We are afraid @Railway has been compromised; they won't tell us.
We have received emails from Anthropic and Stripe that our keys have been compromised.
We are digging into it and will get back to you ASAP!
Durga Sahoo retweeted
I think I found a planet nobody knew existed.
It's been hiding in NASA telescope data for seven years.
At the end of this thread, you can fly to it yourself.
Durga Sahoo retweeted
We just raised $7.7M to kill every subscription for agents.
Introducing monid.ai, the openrouter for agent tools.
Your agents can now:
> discover, run, and pay for tools at runtime
> use 2,500 APIs through one connection: leads, SEO, marketing, search, ecommerce, stocks, video/image/music/3D gen, agent email & phone
> pay per call, 0 subscriptions
Durga Sahoo retweeted
so you didn't go to an ivy league, have a chip on your shoulder, aren't satisfied with your current role and want to work in applied AI
here's what i'd recommend you do immediately
1) get on X and LinkedIn (if you're seeing this congrats)
2) clean up that profile pic (get a headshot srs, use ChatGPT Images)
3) fix your profile, headline and bio (emulate mine if you need a reference)
4) start posting daily about your applied ai learnings, experiments, commentary
what you're trying to do is clear the minimum bar that your profile is actively contributing to the applied ai space. you don't need 5000+ followers. you don't even need engagement.
all you're trying to do is not look like a bot and look like somebody who actually thinks about and works on this stuff.
once you have this, start searching for hiring managers posting about new roles they are hiring for.
you're going to want to tune your feed so these come in as quickly as possible after they post. before your feed is tuned, just actively search for them. use grok bot or muse or codex to do this.
understand this - hiring managers, especially new hiring managers, want to pick the right person. as much as a recruiter makes their lives easier, it also takes them out of their own hiring and filtering process. them posting means they are actively engaged in the search.
this is your opportunity.
like the post and reply that you are DMing them. KEEP IT SHORT.
this does two things:
1) creates two notifications for them, for the like and for the comment
2) creates a pull for them to check the DM or click on your profile
DO NOT overwhelm them with a long message like "I'm so interested! I have 5 years of experience in program management, I just graduated from X university, etc etc."
first of all, it’s thirsty. second of all, nothing you say in that comment will be enough for them to pique interest unless you went to a top school or work at a top company (and if you did, why so thirsty?) so don't do it. stand out with brevity and restraint.
your goal is to hook them to see your profile, your post feed, your pinned/featured post(s). congrats, you just bypassed the application/resume blackhole and got them to see a much larger slice of how you think.
if you've never done this, do it. i guarantee you it will transform your job search.
Durga Sahoo retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Durga Sahoo retweeted
We gave Jev 2,029 real phone calls.
No transcripts or audio; it never heard a word. Our AI receptionist's calls were reduced to pure structure, meaning turns, tool calls, workflow stages and timing.
During the calls, Jev made 38,012 turn-level forecasts at 118 ms median latency, reviewed every call with five typed questions and produced 10,145 answers in 26 seconds with 256 requests in flight.
The experiment was zero-shot, with no fine-tuning or examples from our data. We compared Jev's forecasts with what actually happened in the EHR.
By the halfway point, Jev could meaningfully separate calls that would book from those that wouldn't (AUC 0.78), and near the end it ranked them correctly 94% of the time.
Even though Jev over-focused on visible errors our agent usually overcomes, it's still pretty incredible that it analyzed thousands of real calls in seconds for only $3.
Durga Sahoo retweeted
Base44's ARR is now 2.5x what Wix paid upfront.
That $80M bought an AI app builder in June 2025, and ARR went from $100M to $200M in about five months this year.
The usual knock on AI app builders is that the good projects outgrow them. Once an app gets real users, the code moves to GitHub and a full editor, and the builder drops out of the story. That's been the ceiling on the whole category so far.
Today's launch is an editor, the next step after the app builder and the cloud agents, and it reads to me like a direct shot at that ceiling. Base Code opens any GitHub repo in the browser, including apps that started on other tools, so the builder can stay with the project after the code moves.
Developers build years of habit into their editor, and I'd guess winning some of their daily editing inside a browser tab takes longer than getting a non-coder to ship a first app.
PMs and support folks who work right next to the codebase look like the natural first users. A PM who can read the real repo, even casually, tends to write sharper specs and ask better questions in planning.
App builders are moving into the repo now, and I think the people around the engineers will make it a daily habit well before the engineers themselves.
Try Base Code at base44.com/base-code
🤝 Paid partnership
Today we’re introducing Base Code.
It’s an early preview of how we think software will be built in the future.
It’s a product that encapsulates everything we’ve learnt from:
- How our users are building software.
- How we’re building internally, scaling to hundreds of millions of dollars in revenue while keeping our engineering team very small and focused.
Here are some of the principles behind it, which align with how we at @Base44 think about the future of software engineering:
Cloud:
- Software will move past local desktop environments (where current tools are widely used) and move to the cloud.
- Moving everything to the cloud is not easy. Setting up dev environments with databases, infra components, services, mock data, etc. is easier said than done.
- Base Code first scans your code repo and sets up everything - every infra component (databases, Redis, etc.) - in the cloud. It runs nonstop until your preview environment is ready.
- From an enterprise standpoint, it’s also the logical thing to do: a centralized, governed dev environment instead of handing out keys and secrets to all team members.
- Once Base Code does that, EVERY team member can work in this environment. No more syncing local environments.
Internally, this (cloud) is one of the main things that enabled us to move so fast.
Collaboration:
- Once everything is in the cloud, everybody can write software from anywhere (any browser, your phone, WhatsApp / iMessage coming soon).
- You can easily see what everyone is working on, where they’re at, and how they’re prompting - and jump to their environment in one click.
- We’ve built many great collaboration features from the ground up.
Loops, automations, software factories:
- A full, working cloud environment allows for many advanced capabilities we will unveil soon-think agents running in the browser and testing on every device and in any browser.
- It also allows the use of automations to build software factories-e.g., agents reading support tickets, identifying bugs to fix or features to develop, implementing the changes, verifying them in the cloud, and potentially pushing them.
And lastly, there’s a real advantage in being model agnostic.
--------------
As always, we’re releasing it very early. It’s far from perfect-but we’re looking for early feedback so we can build it together with our great community and make this vision a reality.
*We’re giving it away for free for 30 days*.
Give Base Code a try. Connect any GitHub repo, get your live preview, and share it with your team.
I’d love to hear what works and what doesn’t:
app.base44.com/base-code
Durga Sahoo retweeted
Claude Code tip: once Opus 5.5 is your main model, stop leaving Fable 5.1 sitting idle
put it on call with /advisor
run /advisor fable
Opus 5.5 keeps writing the code
Fable 5.1 reads the full session, every tool call included, and only speaks up at three points:
→ before a plan: is this the right approach?
→ when the same error comes back: am I digging in the wrong place?
→ before "done": what did I miss?
Fable 5.1 reviews. Opus 5.5 ships
Jev engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to Jev in under half a second, and the big model only sees the ones that split
- the full tree
> Opus 5.5 on high runs the main session
> explorer reads the code
> worker edits and runs tests
> researcher pulls the docs
> all three on medium
> Fable 5.1 on call as the advisor
paste the tree and this prompt into Claude Code ↓
"Rebuild my Claude Code setup around this tree:
1. Check ~/.claude/agents and .claude/agents for subagents that already fit explorer, worker and researcher.
> Draft new ones only for missing roles
> Give each model: opus, effort: medium
> Skip any that pin a different model and list them
2. Set the main session to high via effortLevel in ~/.claude/settings.json, and set advisorModel to fable
3. Find anything that keeps the advisor off (CLAUDE_CODE_DISABLE_ADVISOR_TOOL, DISABLE_TELEMETRY, any variable that stops feature-flag fetching) plus CLAUDE_CODE_EFFORT_LEVEL, which overrides subagent effort. Report them, change nothing
4. Add one rule to ~/.claude/CLAUDE.md: consult the advisor before a large plan, when an error repeats, and before calling a long task done
Show me every change as a diff first. No edits until I say go."
↳ code.claude.com/docs/en/advi…
Durga Sahoo retweeted
Claude can now help you build evaluations and hillclimb on them.
In this article, we share guidance on eval design & skills that Claude Code can use to improve your applications.
claude.dev/blog/automating-e…
Durga Sahoo retweeted
建议所有试图走出低谷或者想彻底换一种活法的朋友, 把 Dan Koe 刚发出的这套长文系统连看三遍。
很多人都有过同一种令人窒息的经历: 下定决心要早起、健身、做副业, 结果两到四周之后一切又悄无声息地滑回了原点, 甚至被同一个恶习生生困住整整十年。
绝大多数人以为这是自己意志力不够, 但从物理学上看, 你的起步策略从第一天就选错了。
火箭升空需要把总重量 85% 到 90% 的燃料在短短几分钟内全部烧光, 才能挣脱地表引力; 但只要冲破大气层进入既定轨道, 后续飞行的燃料消耗几乎就是零。
启动你人生的燃料, 和维持你人生的系统, 根本就不是同一种东西。
很多人最大的误区是用温吞的微习惯去对抗十年的重力惯性。
在最脆弱的启动阶段, 你需要的根本不是心平气和的自律, 而是极度浓烈的战术压力。
对自己所处现状的极度厌恶, 或者是主动把自己逼进不可退缩的绝境, 才是能点燃 90% 推进剂的脏燃料。
真正让我觉得有意思的是他指出的高手真相。
顶级创业者和运动员每天工作 16 个小时不觉得痛苦, 不是因为他们每天都在咬牙坚持, 而是因为卓越早就成了他们的出厂设置, 让他们去过平庸的生活反而像下地狱一样难受。
一旦火箭被推上了轨道, 就要立刻丢掉脏燃料, 切换到极简的环境自动回路。
输入端必须戒掉智力肥胖, 把每天刷几百条无用碎片的习惯直接掐死;
输出端则把行为拆解成最小可识别动作, 进健身房只承诺做一组硬拉, 坐到桌前只承诺写一行文字, 并在行为发生后立即绑定确定性的奖励。
痛苦和强厌恶负责把火箭暴力推离地面, 极简环境和最小动作负责让系统在轨道上零阻力滑行。
回看你过去经历过的最好的一次蜕变, 当时把你推离地面的到底是一次具体的暴击, 还是慢慢积累的顿悟?
我先说, 以前逼我彻底戒掉熬夜拖延的, 就是一次因为精力涣散当场丢掉核心客户的惨痛教训。
输入端节食与输出端四步回路的完整操作清单, 我整理好沉在一楼了。
🤖 Made with AI
Durga Sahoo retweeted
Learn to fine tune a model this year
Learn to fine tune a model this year
Learn to fine tune a model this year
Learn to fine tune a model this year
Learn to fine tune a model this year
Start small
Start slow
But make sure you start
Durga Sahoo retweeted
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Durga Sahoo retweeted
Please please please, before you have your grokbot, hermes, or Claude do anything, ask it to go watch 20 of the top videos on YouTube about the topic.
Use this skill:
"I want you to learn the skill of [INSERT TASK], what I need you to do is go on YouTube, and watch the top 20 videos on that topic in the last 6 months. Export the transcripts, and consolidate all your learnings into one skill. If a certain point gets repeated often across multiple videos, give these points more weight in your skill."
I've personally done this with virtually every agent skill I have, be it landing page design, LinkedIn formats, SEO articles...
It takes 20 minutes to most agents to go through everything and will 10x the quality of your agent's output.
Durga Sahoo retweeted
Shopify billionaire co-founder & CEO Tobi Lütke — one of the most analytically rigorous founders alive — uses affirmations.
His exact words:
"If you tell yourself or write down something about yourself 100 times, it writes into the neurofrontal cortex at such a deep level that your brain will start reconciling you to that. It just works."
His personal example:
He was terrified of public speaking.
For one week he sat down every day for 10 minutes and wrote: "I love public speaking."
A week later the fear was gone.
"It's not like a placebo. You just actively change your neurofrontal cortex in this moment."
Your brain starts shaping your behavior around the identity you rehearse most.
Durga Sahoo retweeted
Introducing the Content Machine!
This was the first time I walked through the mechanics of our anti-slop content system & how we drive 10,000,000+ impressions at @tenex_labs with a marketing team of...2.
Thanks to @clairevo for having me on her show to share.
Full-writeup & interview below...
What is the Content Machine?
A directory of daisy-chained skills that turn what you already say into publish-ready content. This is the key way to avoid turning into a slop cannon.
It mines the places you already talk (Slack, Notion, Gmail, Linear, GitHub) plus what the internet is saying, finds the ideas worth writing, interviews you to extract the story, drafts it, and edits it to a 9/10 bar before you post.
What are the principles of this system?
1) It is not a "write me a post" prompt. The core belief baked into it: the raw material must come from you.
2) The machine never invents your voice and never fabricates your insight. It does the research, the structure, and the editing, so your time goes only to the part only you can do. It focuses the human on the first & final mile of the content process.
How is it structured?
A two-layer split between the process & the person
- The process layer is generic and shared: the pipeline, the content-type specs, the copywriting references, the onboarding flow. That is what lives in git and what gets shipped to other people.
- The personal layer is yours and never leaves your machine: content-machine.config.md, creators/<you>/ (profile, style guide, lessons), projects/, published/, oracle-reports/. All gitignored, and the desktop build script refuses to package any of it.
What is the 10-step pipeline?
1) Creator Select. Multi-creator by design. It figures out who this run is for and loads their profile, style guide, and content lessons. Everything downstream is scoped to that person.
- Onboarding (first run only). Scaffolds the workspace, auto-creates your Notion Vault, connects your sources, and builds your voice guide one of three ways: import a guide you already have, feed it writing samples, or sit for a short voice interview.
2) The Oracle. Two idea engines running in parallel:
- Oracle scans what you wrote in the last 7 days across Slack, Notion, Gmail, Linear, and Git, hunting for "spikes," moments where you naturally said something worth expanding.
- The Internet Reader scans what the world is saying, pulled only from the source list in your profile (handles, labs, outlets, keyword watchlist), plus a social sweep across Reddit, X, YouTube, Hacker News, and more (thanks /last30days & @mvanhorn).
- Every idea is scored 0 to 10 (POV strength 25%, story potential 25%, emotional intensity 20%, lesson/framework 20%, depth 10%). Everything qualifying gets written to The Vault, a Notion database that is the durable idea bank.
2.5) Research. Before you get interviewed, a research agent builds a sourced brief:
- key facts with links
- current developments
- what has already been said in-market
- contrarian angles
- open questions only you can answer.
3) Interview Panel.
Six interviewer personas (Ferriss, Rogan, Larry King, Stern, Barbaro, Barbara Walters) ask you one question at a time, each chasing a different dimension: tactics, story, core truth, the hidden thing, clarity, emotional depth.
It pushes back on vague answers and will not advance until it has 2 to 3 specific stories with real details.
4) Production. Your interview becomes a raw markdown file: stories, core insights, quotable moments, the emotional anchor, surprising reveals, the "so what." Your exact words are preserved. This file is the source of truth for everything that follows.
5) Refinement. Now it drafts, and only now. It must read your style guide, your content lessons, and the spec for the chosen format (LinkedIn post, X thread, long post, playbook, podcast promo, reaction post, article, and so on).
6) Writer's Council. Six reviewers score the draft:
Morgan Housel (will this matter in 10 years), Tim Urban (is it confusing), Shaan Puri (would I stop scrolling, plus three alternate hooks), Greg Isenberg (what can someone steal), David Perell (is it personal, observational, playful), and a Slop Detector hunting AI tells.
Each gives what's working, what needs work, a fix, and a score.
7) Revision Loop. Under 9/10 goes back around.
The smart part: fixes get sorted into editorial (the machine rewrites it itself) and information gaps (only you have the answer), and information gaps route back to the Interview Panel with targeted questions rather than letting the machine make something up.
Max 3 editorial cycles. At 9/10 the piece becomes the anchor.
8) Repurposing Engine. One anchor fans out into 10+ derivatives: X article, LinkedIn article, short X posts, short LinkedIn posts, a playbook if there is a framework in it.
Each one is written native to its platform with a fresh hook, not cross-posted, and each runs the full council and revision loop to 9/10 on its own. This is the multiplier.
9) Distribution (optional, off by default). UTM tagging, a scheduled publishing queue, CRM capture of every touchpoint, attribution reporting back to pipeline, and marking the Vault row as Published.
10) The Learning Loop, always running.
After you approve a piece it diffs your first draft against the final, extracts the pattern, and asks you to confirm it.
Confirmed lessons go into content-lessons.md and override the style guide. Once a lesson proves out across a few projects it graduates into the style guide itself.
Your first drafts get better over time instead of you re-explaining preferences.
P.S. i'm thinking about opensourcing this. should i do it?