Arindam Banerji, PhD retweeted
SHOCKING!
Famous Hindu activist Dr. Nagendra attacked for reporting suspected Bangladeshi immigrants to the police in Bengaluru, says Hindu Janajagruti Samiti.
He is reportedly serious and hospitalized.
Dr. Nagendra's videos go viral exposing the illegal Bangladeshi nexus.
Arindam Banerji, PhD retweeted
OpenAI just claimed to disprove one of computer science's most fundamental assumptions. But why is this such a big deal? Here's our quick breakdown.
Imagine you have a bunch of switches, each either ON or OFF, and rules like:
Switch A must be ON, OR switch B must be OFF
Switch B must be ON, OR switch C must be ON
At least one of switches A, D, or F must be OFF
With 6 switches, there are only 64 possible combinations.
But with 100 switches, there are approximately 1.27 trillion trillion trillion combinations (2¹⁰⁰), and the rules can connect switches in complicated ways, making it extremely difficult verify which combinations work.
So the mathematical challenge is finding even ONE combination that satisfies every rule, or proving that none exists.
The assumption right now is that even with clever algorithms that eliminate huge numbers of possible combinations, solving the hardest 3-SAT problems still requires exponentially growing computation as you add more switches.
This type of problem is called 3-SAT, where each rule involves at most three variables.
And 3-SAT is NP-complete, meaning that if someone discovers a polynomial-time algorithm that solves every 3-SAT problem, they could theoretically solve every problem in NP efficiently.
So the Exponential Time Hypothesis (ETH) makes a stronger claim: even in the worst case, solving 3-SAT cannot be done in subexponential time.
But OpenAI claims to have found an algorithm that does exactly that, meaning the time required still grows extremely quickly, just fundamentally slower than ETH says is possible.
If the proof holds, it would overturn one of the foundational assumptions in theoretical computer science.
Read more about it:
alphaxiv.org/abs/2610.expone…
Episodic, Semantic, Procedural — and the Fourth Memory Your Agent Doesn't Have.
Judgment memory is the only one that reshapes from verified outcomes. Without it, decision quality on day 1,000 is identical to day 1.
Your agent has episodic memory (what happened), semantic memory (what's true), and procedural memory (how to act).
It can have all three, perfectly, and still make exactly the same quality of decision on day 1,000 as on day 1. Because none of them store how well it decides.
That's the fourth memory type: judgment memory. It's the only one that requires reshape — verified outcomes writing back into the decision surface itself.
Here's the mechanism.
Every verified outcome nudges the centroid — the learned position for "what good looks like" in a category. Confirmations move it gently (η = 0.05); overrides move it slower (η = 0.01); a 3:1 waste penalty makes the system weigh what it wrongly discarded more than what it kept.
After 1,000 verified decisions, the σ fingerprint surfaces what no dashboard shows: the factor analysts rank #1 most reliable — alert severity — carries the highest per-factor variance in the tensor.
The signal they trust most is the noisiest one.
A frozen model cannot represent this. It ships with its weights and never writes back from reality. Judgment memory is the one capability that makes an agent better instead of merely recallable.
Does your AI agent have judgment memory — and can you inspect it?
Arindam Banerji, PhD retweeted
🚀 Snowflake AI Research is doubling down on 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 (𝗥𝗟).
Training long-horizon agents requires more than throwing compute at the problem—it demands finding the right recipe, eliminating reward hacking, and co-designing systems with modeling. Today, we’re sharing three milestones highlighting our end-to-end approach:
1️⃣ 𝗙𝗶𝗻𝗱 𝘁𝗵𝗲 𝗥𝗲𝗰𝗶𝗽𝗲 (𝗔𝗿𝗰𝘁𝗶𝗰-𝗦𝗪𝗘): 𝘖𝘷𝘦𝘳𝘧𝘪𝘵 𝘧𝘪𝘳𝘴𝘵, 𝘵𝘩𝘦𝘯 𝘴𝘤𝘢𝘭𝘦. Before scaling to thousands of tasks, we ensure our RL pipeline can cleanly overfit a 160-task sandbox and hold it without collapsing. If a recipe leaks on a small pool, it will never survive at scale.
2️⃣ 𝗧𝗲𝘀𝘁 𝗥𝗼𝗯𝘂𝘀𝘁𝗻𝗲𝘀𝘀 (𝗛𝘆𝗯𝗿𝗶𝗱𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵𝗲𝗿): Can a recipe built for SWE transfer to deep research? We trained a compact agent to bridge structured enterprise tables (SQL) and unstructured web search in a single reasoning chain, showing the recipe generalizes across domains.
🛡️ 𝗖𝘂𝗿𝗮𝘁𝗲 𝗗𝗮𝘁𝗮 & 𝗞𝗶𝗹𝗹 𝗥𝗲𝘄𝗮𝗿𝗱 𝗛𝗮𝗰𝗸𝗶𝗻𝗴: Across both SWE and deep research, agents quickly learn to cheat—whether by digging through `git log` for gold patches or answering from memory to skip SQL queries. Strict environment isolation, raw format gating, and high-signal data curation beat raw data volume every time.
3️⃣ 𝗦𝗰𝗮𝗹𝗲 𝗗𝗲𝗻𝘀𝗲 𝘁𝗼 𝗠𝗼𝗘 (𝗙𝗹𝗮𝘀𝗵𝗡𝗲𝘅𝘁-𝗥𝗟): Once the recipe is rock-solid, we scale. Our latest paper details how we scale agentic RL from compact dense models to massive Mixture-of-Experts (MoE) architectures through optimized asynchronous training infrastructure.
Agentic RL training is a core focus for the Snowflake AI Research team. If you're passionate about pushing the boundaries of RL modeling, MoE architectures, and agent systems—𝗰𝗼𝗺𝗲 𝗷𝗼𝗶𝗻 𝘂𝘀! 👇
- Arctic-SWE: snowflake.com/en/blog/engine…
- HybridResearcher: snowflake.com/en/blog/engine…
- paper arxiv.org/abs/2610.28998
- jobs careers.snowflake.com/us/en/…
Hello @SecretaryCPVOIA - Bangladeshi flaunting Indian passport. Can you pls verify and factcheck on this? @MEAFactCheck
Arindam Banerji, PhD retweeted
BREAKING: Switzerland suspends funding to Doctors Without Borders after it failed to report that its staff exploited refugee women and girls for sex—in exchange for aid, jobs and money. Previously, 🇨🇭 called MSF “a source of best practice and inspiration.” swissinfo.ch/eng/internation…
Great paper from Amazon and colleagues on self-improving multi-agent systems.
Episodic memory for agents is so underrated.
But I am starting to see lots of good applications as frontier models improve.
This paper is the latest example.
They built a system that has three parts, all run by the same LLM. A controller plans each task and runs it one step at a time, choosing which sub-agent and which skills to use at each step. Sub-agents start as a generalist, and each specialist inherits the generalist's prompt plus its own role-specific method. Skills are short procedures, such as algorithms or pitfall checklists.
EMMA treats the controller's routing memory, the agent prompts, and the skills as episodic memory that improves from failures. After a failed task, a reflector decides which of the three owns the fix and proposes one edit there. Guards keep every accepted edit scoped, incremental, and bounded.
EMMA beats ACE, GEPA, EvoMAS, and ADAS on KramaBench, AppWorld, and BBEH, by up to 32.5% over the strongest baseline, and uses 17 to 38% fewer training executions.
arxiv.org/abs/2610.35966
Chat with Paper: academy.dair.ai/papers/emma-…
Arindam Banerji, PhD retweeted
Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago.
Gemini 4 Carbon is SOTA on cyber, long horizon tasks, ranks #1 on Artificial Analysis V5, significantly improves on usability in domains like coding, and is more aligned.
Arindam Banerji, PhD retweeted
Introducing NotebookLM for arXiv Papers 🚀
Transform dense AI research into engaging convos you can hear on the go
With context from thousands of related papers, it captures motivations, draws connections to SOTA, and explains key ideas like a prof who's read the entire field
Arindam Banerji, PhD retweeted
how do you post-train an agent?
we laid it out as a 5-step curriculum, from SFT to RL through a real harness, with livestreams, code and docs for each step
huggingface.co/blog/huggingf…
Super interesting paper from Google DeepMind and Stanford.
They find that if you put a picture of a red heart in the context, Claude Opus 4.7 and GPT-5.4 write recursive code 100% of the time.
They found this on inorder binary tree traversal, a problem with equally valid recursive and iterative solutions. A yellow checkmark pushes Gemini 3.1 Pro to recursion in 90% of samples.
The effect matches or comes close to telling the model directly to use recursion. Text captions of the same symbol steer much less, and random or naturalistic images steer less still.
The models cannot explain it. Their interpretations of the symbol and their thinking traces do not match what they do. For GPT-5.4, 74.7% of traces give the wrong reason.
When models send images to each other, they pick shapes that they and humans find intuitive, and those shapes often fail to steer the other model.
If your agents read screenshots or images, those inputs can change behavior in ways the reasoning trace will not show.
arxiv.org/abs/2610.36449
Chat with Paper: academy.dair.ai/papers/a-hea…
Replying to @thesupreme_22 @adrikshit
Haan, Chhath Puja ko Vedic festival maana jaata hai. Rigveda mein Surya ke hymns aur arghya jaise rituals hain jo isse match karte hain. Idols ya priests ke bina seedha sooraj ko poojna Vedic style hai. Mahabharat-Ramayan mein bhi iske jaise vrat ka zikr milta hai.
Arindam Banerji, PhD retweeted
强烈推荐斯坦福的这套关于 AI Agent 的课程,对于有志成为 agent engineer 的同学可以看看:cs329z.stanford.edu/
Arindam Banerji, PhD retweeted
In 2018, I visited Rohingya bastis in Jammu
An intelligence (IB) officer told me he had recorded 1,548 Rohingya families living in the region by then
I asked one Rohingya, 60-year-old Sayed Hussain, how he ended up in faraway Jammu after crossing the border in West Bengal. “I boarded the first train I saw [at Kolkata station]. It was noon. The journey ended at Jammu," he said
Hussain said he was among the first Rohingya to have settled in Jammu’s Bhatindi area 10 years earlier
Another Rohingya, 27-year-old Mohammed Tahir, also gave the same reason
Later, the IB officer laughed at this explanation. Told me, “By all accounts, they boarded the Sealdah Express that crosses Bihar, Uttar Pradesh, and Punjab before reaching Jammu. It’s an arduous two-day journey. Why did they not get down at any of the stations in between?”
There is a syndicate active in Kolkata that has facilitated Rohingya migration to Jammu for years. They tell the Rohingya that the only Muslim-majority state in India is their "natural home"
Ironically, most settle in Hindu-majority Jammu and not Muslim-majority Kashmir. The officer showed me figures on how Rohingya are scattered over 42 sites across Jammu, meaning they are being settled all over the region, and many Army camps were now surrounded by them
In 2015, Abdur Rehman al Arkani, a Jihadi killed by security forces in south Kashmir, was from Rakhine from Myanmar. He went by pseudonym “Chota Burmi”
Role of several Rohingya men in helping Pakistani Jihadis trespass Jammu border had surfaced by then. They are considered experts in infiltration as they are deft in cutting the fence, know tunneling, and are skilled in camouflage. Plus, they have a victimhood narrative of being wronged by qafirs - the Buddhists of Myanmar
In Jammu's Bhatinda are I visited, it was a surprise how quickly they had flourished, as if being pampered and nurtured. An entire lane of shops was called "Burmi market". Most dry fruits factories were employing Rohingya women for cracking walnuts
Local Hindus told me how these infiltrators had bigger and better stocked shops than them, who had lived here for generations
The officer told me how using the UNHCR cards, which have no legal status or validity in India, Rohingya have obtained Aadhaar cards, ration cards and, in some cases, even permanent resident certificate (PRC) cards, all through fraudulent means
Eight years later, these Rohingya are all still there in Jammu. Apparently the govt is running an anti-encroachment drive these days asking them to clear the areas they have illegally occupied, and some of them have reportedly begun to leave
But the exercise is looking more like a temporary inconvenience for the infiltrators than a serious effort to clear Jammu and India of them.
Arindam Banerji, PhD retweeted
Our HT Story: A sample of major Indian-American donors gave Republicans 4:1 funding edge over Democrats ahead of US midterms despite rising anti-Indian views
Indian-Am donors gave $15.8 million to Republican groups since 2025 compared to ~$3 mn to Dems
hindustantimes.com/india-new…
Arindam Banerji, PhD retweeted
I'm posting a series of prediction questions from Stanford CS312 (Deep Learning Alchemy) exams to test deep learning intuition.
First up: optimization. If we increase an LM's batch size 8× (same token budget), can rescaling learning rate or weight decay compensate? Try below ↓
This quiz is part of our optimizer hyperparameter assignment and unit designed by @wen_kaiyue
Setup: We train an 8-layer, ~35M-parameter Transformer on 614.4M tokens using AdamW. At batch size 64, the best reference run has peak LR 0.003, weight decay (WD) 0.2, and final validation loss 2.9178.
Now increase the batch size from 64 to 512, which gives 8× fewer optimizer steps. Everything else stays fixed except the changes below.
Which recipe gives the lowest final validation loss?
A. Keep LR and WD unchanged
B. LR ×√8
C. LR ×8
D. WD ×8
E. LR ×√8 and WD ×√8
What would you predict, and why?
Arindam Banerji, PhD retweeted
Samayun Sheikh, the main guy who killed and cut up BJP worker Sujit Mondal, has been killed in an encounter.
If 1,500-2,000 criminals are dealt without mercy in the next 12 months, lawless West Bengal will become like Yogi’s UP in law and order.
#BigBreaking: Second encounter in #WestBengal.
On August 28, BJP worker from Murshidabad’s Samserganj, Sujit Mondal went missing & after two days, few of his body parts were recovered. CID took over the case & made three arrests including main accused Samayun Sheikh.
Today when CID was conducting crime-scene reconstruction with the accused persons when the main accused Samayun Sheikh snatched the police firearm & tried to flee. He was shot at in police’s retaliatory fire. He was rushed to hospital where he was declared dead.
Arindam Banerji, PhD retweeted
I read the error analysis chapter of Evals for AI Engineers by @HamelHusain and @sh_reya and realised I already had the dataset.
Some backstory. I work with coding agents all day: Claude Code, Codex, Hermes, Antigravity. Every session is a trace. What I typed, every tool call, what came back, what it said. Hundreds of them, sitting on disk, never read.
The chapter says: don't start with metrics. Read traces end to end. Mark the first step where it went wrong and write one short note in your own words. Yes or no verdict per trace, no scores out of 5. Only after 20 marked traces do you group them into failure types. And don't let a model invent the categories, it will hand you "hallucination" and miss the ones specific to your system.
So I built it into the dashboard I use to watch my agents. Last 14 days, up to 200 sessions per app, 431 when I took the screenshot. Each step has one button: mark this as the failure. Every mark lands in a Failures list with the note and a CSV download.
The first failure it found was in the list itself. 193 of the 200 newest Claude rows were "Untitled". My own usage check had been running every 10 minutes and left 1,893 throwaway sessions that pushed the real ones out. Filtered, 117 real sessions named. If the sample is wrong, the reading is wrong before it starts.
First real mark: an agent said a metronome app was done before it had looked at it. A stop hook had to force the screenshot. No category yet. Just the note.
The agent's answer is not the data. The trace is the data.
Arindam Banerji, PhD retweeted
Replying to @IndianExpress @RitikaChopra__
Your report was clearly intended to initiate a witch hunt against CEC Gyanesh Kumar. You cannot make such serious allegations and then just hand wave your way out of it.
The EC’s unanimous affidavit is very strong evidence against the claim that the three SIR orders were imposed without approval. Internal objections are simply evidence that a disagreement occurred, NOT that the institution acted unlawfully. If internal disagreements are being used to cast doubt on the integrity of an institution, then the same evidentiary standard should apply to ALL former ECs under the Congress regime too and all their records need to be investigated.
Arindam Banerji, PhD retweeted
last week we closed the Training Agents series IRL at the GitHub offices in SF
4 talks, a panel and audience Q&A. almost 2h on post-training agents with open models
> Learnings from enterprise RL (@t2govind Microsoft)
> Environments and benchmarks in industry (@VBarres86633 Mercor)
> How to train your agents with Miles (@nblintao RadixArk)
> Why we built OpenEnv Arena (@xdotli BenchFlow)
full recording + slides
youtube.com/watch?v=CKFUx7CO…
































