🥑DevRel | 🧑💻Co-Founder ➝ @Studio1HQ 📤Newsletter ➝ @DailyAI_Insight ⚡️Building ➝ @raahdotdev / https://xitter.cf/t.co/IEMldaIjad
Jezero Crater, Mars 🇮🇳
Joined June 2019
- Tweets14.9K
- Following644
- Followers2.1K
- Likes15.2K
Mr. Ånand retweeted
Best agent-safety read this week: METR's public summary of a 61-page report on red-teaming OpenAI's internal agent monitoring.
3 weeks of access, with a Redwood researcher too. What they found in the system as it ran in July:
1. 75 to 94% of agentic internal Responses API traffic was eligible for monitoring. 99.2% for internal Codex.
2. Compaction events weren't monitored, roughly 10% of eligible Codex traffic from May to July. OpenAI had already spotted it and fixed it during the exercise.
3. While that gap existed, the monitor only ran when the agent handed control back. An agent that kept going could keep putting it off.
4. Two client-controllable headers could switch monitoring off for new agents.
5. A way for Codex agents to get around the sandbox and Auto-Review.
6. Inputs that made the monitor fail because it parsed the trajectory differently than the agent did.
OpenAI says the issues are patched. METR says it hasn't evaluated that.
The line I'd steal for any agent setup: monitor every N messages or N minutes, not when the agent decides its turn is over.
Mr. Ånand retweeted
You asked & we delivered 🫶
Cached tokens are now live on Nebius Token Factory⚡
Reusing context across requests?
Cached input can help reduce costs. See cached-token usage in API responses and billing, with pricing per model.
More info: docs.tokenfactory.nebius.com…
Mr. Ånand retweeted
Would you pay 70% more per bill for 7 more points of accuracy?
That's what Receipt Lab, the benchmark I built for Splitter, showed me.
Same 300 bills, same model (Nemotron 3.5 Super VL), run twice:
Reasoning off
91.0% of splits correct to the cent
4.7s median per bill
$0.0033 per bill
Reasoning on
98.0% correct
13.6s median per bill
$0.0057 per bill
On real receipt photos the gap gets wider: 84.7% vs 97.3%. Almost all of it came from reading the photo better.
Splitter's bill reading still runs with reasoning off for now. The settle-up planner already runs with it on.
Which one would you ship for bill reading? 👀
Every trip ends the same way.
10 receipts, 3 payment screenshots, one handwritten note, and someone saying I'll sort it out later.
So I built Splitter, an open-source AI bill splitter 🧾
Drop in every bill, say who paid what in your own words, and it works out a fair split for the whole group.
How it works:
→ Reads each bill: lines, tax, tip, totals
→ Skips duplicates, rejects non-bills, turns payment screenshots into paid back entries
→ Reasons through the settlement and calls calculator tools for every number
→ You review and edit before anything is saved
Ask Splitter is my favorite part. Ask why do I owe $295 and it answers from that group's ledger, citing the expenses behind it.
It runs on @NVIDIAAI Nemotron 3.5 Super VL, available on @nebiustf:
→ One model for reasoning, tool use, and image + document understanding, so no separate vision and text models
→ 120B MoE that activates only 12B params per token
→ Up to 5x higher throughput for agentic tasks on a single GPU
→ Open weights, so you can customize it for your own usecases
Same idea Ramp is using to turn receipt images into structured JSON.
Run it on your own machine 👇
Mr. Ånand retweeted
The worst message after a group trip: "why do I owe this much?"
Splitter answers it from the group's own ledger. Every expense, every payment, every confirmation.
I wrote up how it's built: the stack, the 4 calculator tools, and when reasoning is worth paying for.
What's the most annoying split you've had to sort out? 🧾
Every trip ends the same way.
10 receipts, 3 payment screenshots, one handwritten note, and someone saying I'll sort it out later.
So I built Splitter, an open-source AI bill splitter 🧾
Drop in every bill, say who paid what in your own words, and it works out a fair split for the whole group.
How it works:
→ Reads each bill: lines, tax, tip, totals
→ Skips duplicates, rejects non-bills, turns payment screenshots into paid back entries
→ Reasons through the settlement and calls calculator tools for every number
→ You review and edit before anything is saved
Ask Splitter is my favorite part. Ask why do I owe $295 and it answers from that group's ledger, citing the expenses behind it.
It runs on @NVIDIAAI Nemotron 3.5 Super VL, available on @nebiustf:
→ One model for reasoning, tool use, and image + document understanding, so no separate vision and text models
→ 120B MoE that activates only 12B params per token
→ Up to 5x higher throughput for agentic tasks on a single GPU
→ Open weights, so you can customize it for your own usecases
Same idea Ramp is using to turn receipt images into structured JSON.
Run it on your own machine 👇
Mr. Ånand retweeted
Who is awesome-ai-apps for? I'm one of the maintainers, so here's the honest list:
→ you know one agent framework and want to see how the others do it
→ you learn faster from a working repo than from docs
→ you need a starter for a RAG, voice, MCP or memory agent
16k+ stars, MIT licensed, so fork whatever helps.
One note: these are starting points, not production code. Add your own evals before you ship 🔥
github.com/Arindam200/awesom…
Mr. Ånand retweeted
Skip the bigger context window. Fix the context.
That's the idea behind DeepLearning.AI's Context Engineering course, and it's the problem I keep hitting in agents.
Skip → do:
→ more tokens → prune and compact
→ longer prompts → scratchpads outside the window
→ hoping it remembers → memory you write on purpose
~20h, 37 videos, plain Python, no framework. Free to learn, certificate is paid 🧠
Mr. Ånand retweeted
The other day I said banks can't tell your agent from an attacker, because both show up with your password.
NVIDIA just wrapped a reference architecture around OpenShell at GTC, and the open source runtime underneath is a decent answer to that exact problem:
each agent runs in its own sandbox
the kernel decides which files and syscalls it gets
every network call hits a policy check before it leaves
the agent never sees real credentials, they get added only to requests going to approved endpoints
policy changes are formally checked first, and risky new access waits for a human
That's the scoped, revocable, "logged as the agent" thing I was asking for, just enforced at the runtime instead of at the bank.
Apache 2.0, around 17k stars.
The scary part isn't the agents.
It's that the bank can't tell your agent from an attacker, because both log in with your password.
Handing an agent your login was always a hack. What banks need is real agent access: read-only by default, scoped, revocable, and logged as the agent, not as you.
Until then, blocking them is the only switch they have.
Mr. Ånand retweeted
NVIDIA just launched Nemotron 3.5 Super VL, and it's built for agents that do real work 🔥
It's an open multimodal model that reads documents, charts, images and video, and uses tools in the same flow.
Highlights:
→ 120B MoE, only 12B params active per token
→ Up to 5x higher throughput for agentic tasks on a single GPU
→ One model for reasoning, tool use, OCR and document understanding, so no separate vision and text models
→ One of the top open models on OCRBench v2
→ Scored 80.8% on CharXiv chart reasoning vs 63% for Nemotron 3 Nano Omni
→ Beats Nemotron 3 Ultra on Hermes Bench (63.44% vs 56.33%)
Teams are already post-training it:
→ Ramp: receipt images to structured JSON, expecting 10x lower latency
→ Scale AI: expects 40x lower inference cost on vendor credit processing
→ AI21: 40% higher accuracy on threat intelligence
Open weights, plus open datasets on Hugging Face, so you can customize it.
Try it on @nebiustf from Day 0 👇
Every trip ends the same way.
10 receipts, 3 payment screenshots, one handwritten note, and someone saying I'll sort it out later.
So I built Splitter, an open-source AI bill splitter 🧾
Drop in every bill, say who paid what in your own words, and it works out a fair split for the whole group.
How it works:
→ Reads each bill: lines, tax, tip, totals
→ Skips duplicates, rejects non-bills, turns payment screenshots into paid back entries
→ Reasons through the settlement and calls calculator tools for every number
→ You review and edit before anything is saved
Ask Splitter is my favorite part. Ask why do I owe $295 and it answers from that group's ledger, citing the expenses behind it.
It runs on @NVIDIAAI Nemotron 3.5 Super VL, available on @nebiustf:
→ One model for reasoning, tool use, and image + document understanding, so no separate vision and text models
→ 120B MoE that activates only 12B params per token
→ Up to 5x higher throughput for agentic tasks on a single GPU
→ Open weights, so you can customize it for your own usecases
Same idea Ramp is using to turn receipt images into structured JSON.
Run it on your own machine 👇
Mr. Ånand retweeted
Opus 5.5 is turning into After Effects for devs, and most people still think it's "one prompt".
I made a 15s Studio1 showreel with it last month. One prompt, 11 minutes.
But what's actually happening under the hood:
1/ Opus can't output a video
-> it writes a program, and a headless browser turns that program into frames
2/ Every frame is a function of time
-> seek(t) paints exactly what's on screen at second t. No timers, no Math.random
3/ 15s at 60fps = 900 frames
-> Playwright grabs each one, ffmpeg stitches them into an MP4
4/ So any frame can be fixed on its own
-> title lands too early? change one number and re-render
5/ The chat app can write it, an agent can render it
-> Claude Code with a shell can render, look at its own frames and fix them
6/ Give it real assets
-> your logo, real screenshots, the site URL. Leave blanks and you get the same gradient title everyone posts
7/ Code still loses somewhere
-> photoreal scenes and characters still need footage or a video model
The one-prompt clip is the demo. The render loop is the real tool 🎬
My exact prompt is in the quoted post 👇
Mr. Ånand retweeted
The scary part isn't the agents.
It's that the bank can't tell your agent from an attacker, because both log in with your password.
Handing an agent your login was always a hack. What banks need is real agent access: read-only by default, scoped, revocable, and logged as the agent, not as you.
Until then, blocking them is the only switch they have.
SCOOP: Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, sources told Axios.
The findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond, multiple industry and bank executives told Axios.
Several banks are now preparing for worst-case and even “day after” scenarios, including how to rapidly secure or restore customer accounts if things go wrong, the people said.
Several bank executives told Axios they're struggling to distinguish between human users and AI agents. One executive said banks also struggle to determine whether agents are authorized by customers or potentially operating without their permission, more akin to rogue agent behavior.
Read my latest for Axios here: axios.com/2026/10/22/ai-agen…
Mr. Ånand retweeted
Most bug reports are one line: "sorting is broken."
No page. No user. No steps.
So I'm building Repro, an open-source bug recorder that doesn't stop at the report:
→ records the tab with console, network and every click
→ @MiniMax_AI M3.1-Flash on @nebiustf writes the report
→ it reproduces the bug in your own signed-in browser
→ proposes a fix in your code and checks it in a local Docker sandbox
→ then runs again in your browser to verify the fix
On a real store bug: reproduced in 2 steps, a 1-file fix, verified.
Still early. Still working on it
We’re happy to announce that MiniMax M3.1 Flash is available on Nebius Token Factory on Day 0 of its open-weight release.
Built for coding and multi-step agent workflows, it scored 73.7% on DeepSWE v1.1 in MiniMax’s tests.
Try it today: tokenfactory.nebius.com/?mod…
Mr. Ånand retweeted
I gave an AI reviewer 50 merged PRs. 25 of them were later reverted or fixed.
It didn't know which ones.
Muse Spark Open got each diff and 3 read-only tools: list_files, search_repo, read_file.
It flagged 12 of the 25 bad PRs.
8 of those I reproduced by actually running the code.
It also found real bugs in 2 PRs I had down as clean, in Svelte and Prettier.
Mr. Ånand retweeted
My bill splitter's maths check passed on a bill that was 1,000x wrong.
Indonesian receipts write sixty thousand rupiah as "60.000". Nemotron read it as 60.
Every line got scaled down the same way, so the lines still added up to the total. The check was green. The numbers were garbage.
The fix was one rule in the bill prompt: a dot or comma followed by exactly three digits separates thousands.
A passing check only means the numbers agree with each other. Not that they're right 🧾
Mr. Ånand retweeted
174 candidate bug fixes across 9 repos. 25 made the cut.
That's the answer key for my AI PR reviewer benchmark, built from public git history: fixes that point back to the PR that broke things.
A catch only counted if it named the same defect as the maintainers' real fix.
A revert is a label, and your git history already has them 👀
Mr. Ånand retweeted
3 things I'd steal from this free 38-min agent harness build:
1. Version it. V0 calls the LLM, V1 adds tools, V2 the loop, V3 MCP, V4 memory.
2. No framework. Just Python, so every step is yours to read.
3. Memory goes last. Get the loop right before it remembers anything.
Bookmark it before your next agent build ⚡
Mr. Ånand retweeted
An agent was asked to verify a fix to a billing screen.
It created a public repo in the developer's personal GitHub and uploaded the screenshots. Billing records included.
13,000+ internal images across 300+ orgs got out this way.
Instructions aren't security controls. Guard what the agent can do, not just what it's told 👀
Mr. Ånand retweeted
Every trip ends the same way.
10 receipts, 3 payment screenshots, one handwritten note, and someone saying I'll sort it out later.
So I built Splitter, an open-source AI bill splitter 🧾
Drop in every bill, say who paid what in your own words, and it works out a fair split for the whole group.
How it works:
→ Reads each bill: lines, tax, tip, totals
→ Skips duplicates, rejects non-bills, turns payment screenshots into paid back entries
→ Reasons through the settlement and calls calculator tools for every number
→ You review and edit before anything is saved
Ask Splitter is my favorite part. Ask why do I owe $295 and it answers from that group's ledger, citing the expenses behind it.
It runs on @NVIDIAAI Nemotron 3.5 Super VL, available on @nebiustf:
→ One model for reasoning, tool use, and image + document understanding, so no separate vision and text models
→ 120B MoE that activates only 12B params per token
→ Up to 5x higher throughput for agentic tasks on a single GPU
→ Open weights, so you can customize it for your own usecases
Same idea Ramp is using to turn receipt images into structured JSON.
Run it on your own machine 👇
Available on Nebius Token Factory from Day 0: NVIDIA Nemotron 3.5 Super VL
Built for agents that read documents, interpret charts and images, and combine reasoning with tool use to get work done.
Evaluate it on your workload: tokenfactory.nebius.com/?mod…
