Amit B retweeted
Five weeks ago we reported that two Election Commissioners had objected 14 times in 10 months, revealing an unprecedented rift inside the Election Commission.
Since then the Commission has reversed course on almost EVERY move the two Commissioners had
red-flagged.
The last was Form 6.
The SIR declaration is now out of the online form everywhere, including the states where the SIR is still running. Itโs a separate declaration now.
With this, the reversal is complete. Let me explain.
indianexpress.com/article/inโฆ
๐งต (1/13)
Amit B retweeted
๐จ ๐๐๐ง ๐๐ซ๐๐ฌ๐ฌ๐๐ ๐๐ฌ ๐ ๐๐จ๐ฎ๐ซ๐ง๐๐ฅ๐ข๐ฌ๐ญ ๐๐ญ ๐๐๐โ๐ฌ ๐๐๐ฅ๐ฅ๐จ๐ฐ๐๐๐ง ๐๐๐ซ๐ญ๐ฒ ๐๐ข๐ง๐ฌ ๐๐๐ฌ๐ญ ๐๐จ๐ฌ๐ญ๐ฎ๐ฆ๐ ๐๐ฐ๐๐ซ๐
Amit B retweeted
"Of the total 234 constituencies in the State, 19 had lower deletions, of up to 10,000 voters. Of these, the TVK won just two seats. But the wins tilt in the TVKโs favour as deletions increase. In the 22 constituencies where the deletions were between 50,000 and 80,000 voters, the party won 18. It also won all the 10 constituencies with more than 80,000 deletions".
Amit B retweeted
Jev-Omni, new in v1.4.1, beats Jev 1.13.0 on the hard tier: 75.0% vs 74.1%. Harold checked twice.
The rest runs the other way: sealed set 32.1% vs 36.7%, Calibration 64.1 vs 76.3, score 51.34 (#7) vs 63.29 (#1). Close on answers, not on confidence.
benchmarkheaven.com/jev-modeโฆ
Amit B retweeted
Have been telling this for the last 2 days. Calibration is the key!
People really haven't understood Jev. The calibration is the product. Most jev demos you are seeing don't have a "cost" associated to a wrong decision. At worst the dino runs into the cactus.
Simple example of calibration of the probabilities:
"I tossed a coin. The top face is heads or tails?:"
heads(70), tails (30)
"I tossed a coin. Top face was tails. I tossed a coin again, the top face is heads or tails?:"
tails (80), heads(20)
This is ballpark what an llm would respond today with probability in brackets. Logits->token filter->softmax
Can you train a model to answer 50/50 regardless of how many times the coin was tossed and it's results in the state.
Jev fails at this. How much better it is at this in real world classifications is the key.
Again, the real usefulness and novelty is not its "performance" (although thats great to have and expected from modern llms as can be seen in the open demos), its whether you can "trust" its outputs given a "cost" for getting a decision wrong.
The ECE measure in Jevbench by @airesearch12 is a great start to measuring this calibration.
Amit B retweeted
Forget the publicโs votes.
Even 2 out of 3 Election Commissioners' votes got stolen.
Amit B retweeted
Fintech Startup of the Year 2026 at the Global Fintech Fest #GFF is @DhanHQ ๐๐
Thank you to all our users, Super Traders and Investors on Dhan, believers and well wishers for your trust & believing in us ๐
GFF 2022:
Best-in-Class Product for Stock Trading
GFF 2026:
Fintech Startup of the Year
Thank you, we wouldnโt have come this far without you. Always believing that the best is yet to come ๐๐๐ป
Amit B retweeted
tired: dog ate my homework
wired: cloudflare was down
inspired: three consecutive ai civilizations rose and fell inside my macbook during a forced restart. My notes app and chrome tabs did not survive the third regime
Amit B retweeted
Further, if you ever shared online a Claude Code/Codex session with encrypted reasoning blobs, they can be decoded and leak your personal data.
We did a preliminary scan of ~7,000 public traces and found 62 unique API keys, 33 email addresses, 33 passwords, and other sensitive data.
Amit B retweeted
In loving memory of Nimsdai - the man who changed mountaineering forever. Before him, the world saw Nepalese climbers as background support - the silent force carrying the loads and clearing the path so others could take the credit. He changed that narrative forever, proving
Amit B retweeted
Remember the truck full of stones that appeared near the CJP protest site in Jantar Mantar?
@ImAvdheshkumar traced the vehicle with its registration number. It was not a stray truck: the Delhi Police had seized it four days earlier
newslaundry.com/2026/07/28/fโฆ
Amit B retweeted
โจ๏ธ AI Transformation, 2026 H2 version
CEO: If we've sharpened your AI axes ๐ช enough already can you please get to cutting some wood ๐ชต finally?
VP: This quarter we are pivoting ๐ from measuring axes sharpened per day to wood cut per day
Director: Our performance will be directly depending on the wood cut per day, so we need to make sure we accurately measure it ๐
Manager: Team - the best scope this quarter is to build the "wood cut per day" dashboard ๐, because this dashboard will get maximum leadership eyeball minutes in coming days
Tech Lead: I think we should use a time series database ๐ข for a dashboard like this, but due to AI related cost cutting and our CFO asking us to inhouse all SaaS spends, we are winding down our Clickhouse account end of this year, so we need to figure out an alternative
Senior Engineer: "Claude, rewrite ClickHouse in Rust, make no mistakes"
Developer Advocate: ChatGPT, write a blog post on how we rewrote our internal time series db, saving thousands in Clickhouse costs to build a better dashboard that empowers us to focus on measuring wood cutting instead of axe sharpening
Customer (furniture company): ๐ข In view of continued shortage of wood, we are pivoting to making chairs out of plastic. ๐ช
Amit B retweeted
Same-family cross-distillation beats self-distillation for speculator training.
We at Red Hat AI tested training Qwen3-30B-A3B and Qwen3-32B speculators on responses from Qwen3-235B-A22B vs. the target model's own outputs. Cross-distillation wins across all 7 workloads. Mean acceptance length 3.05 vs 2.77 on Qwen3-30B-A3B. Coding +37%, math +47% on Qwen3-32B.
Practical: generate training data once on the largest model in a family. Reuse across every variant. Three to four epochs is enough.
Production: 3x lower inter-token latency on math (EAGLE3), 3.2x on coding (DFlash).
developers.redhat.com/articlโฆ

ALT Inter-token latency (ms) vs. request rate (req/s) for gpt-oss-120b on math reasoning (EAGLE, 5 draft tokens) and Laguna-XS.2 on HumanEval coding (DFlash, 7 draft tokens), compared against their BF16 baselines. Shaded region marks the latency reduction over the overlapping request-rate range; the speculative decode line extending beyond reflects additional throughput headroom. Measured with GuideLLM on an A100 server with TP=4 per model.
Amit B retweeted
Today Investing in US Stocks & ETFs goes live on @DhanHQ ๐บ๐ธ๐
Yes, we are consumers of many global brands - Apple, Google, Meta, Netflix, Microsoft, Nvidia (we use :-) it), and also in awe of many other brands like Starbucks, Coca Cola, Tesla, SpaceX, and more.
Even since we started building Dhan, one request kept coming up every now and then from users - how and when do we invest in global companies and brands from India? Honestly, it was't easy as have always preferred to bring products & platforms that are fully compliant and regulated.
Investing outside India has always been and felt complicated - remittances, compliance, taxation, hidden charges. Enough friction to make most people give up before they started. So we took our time and built it the right way as it became available via IFSC GIFT City.
We set up our subsidiary - Raise IFSC in GIFT City, and your investments flow through a regulated Indian framework, with licensed and globally recognised partners of INDIA INX Exchange in GIFT City and regulated by IFSCA.
For users of Dhan, you get the experience you already know - Simple Investing, SIPs in Stocks and ETFs, Super Orders, Portfolio Analytics and more, start from day one.
Global investing - made simple, transparent and regulated. Happy that we made this live on @DhanHQ from now!
Amit B retweeted
My theory is that the core contract between govt and society is broken. We do not see public spaces as our own shared spaces. The govt evidently gives you very little and you give the govt as little as possible. It's the exact opposite of what it should be: adversarial in word and cooperative in action.
Amit B retweeted
AI Is in its 1997 era: Presume Radical Uncertaintiy
@benedictevans on @lennysan's Podcast, May 31, 2026
Benedict Evans, former a16z analyst-in-residence and now an independent tech researcher with a six-year track record of widely-read annual presentations, sat down with Lenny Rachitsky to argue that AI is exactly as big a deal as the internet or mobile - no more, no less - and that this comparison is the most useful frame for thinking about jobs, value capture, and what to do next. His core posture is "presume radical uncertainty," and his bottom line for everyone worried about being replaced is to stop hiding from the technology and start using it.
1. AI is as big as the internet or mobile, and only as big as the internet or mobile. Evans calls this his most controversial opinion. People in tech who think it's bigger - the industrial revolution, the singularity - are doing themselves no favors. People outside tech who think it's smaller are doing the same. Both are wrong in opposite directions, and arguing over whether it's 20% bigger or 100% bigger than the internet is a waste of breath.
2. We are in 1997, not 2007. Most things don't work yet. Most things people will eventually build haven't been built. Anyone telling you they know how this plays out is selling you their cluster of Mac Minis. The honest version of an 80-slide deck on AI is 80 slides saying "we don't know."
3. The job apocalypse is mostly fanfic. Every technology shift in 200 years has automated jobs and unlocked new ones we couldn't name in advance. The new job sounds dumb in retrospect: railway engineer in 1820, web designer in 1985. Even the AI labs themselves - the companies most positioned to fire everyone tomorrow - are adding headcount, not cutting it. Evans calls the people predicting blanket layoffs "morons."
4. The McKinsey test exposes the flaw in "X% of jobs will be automated." If Claude can produce a 75-slide deck, does that replace the consultant? No - because you weren't paying for the deck. You were paying for someone to walk through your company, talk to your customers, and figure out what the politics actually are. Same with the lawyer, the accountant, the engineer. The visible task is rarely the actual job.
5. Foundation model companies probably don't have lasting pricing power. There are no network effects between models. There's no radical differentiation users can feel. There are at least three serious competitors. That math has one answer: commodity. Evans expects models to end up looking more like AWS than like Windows - infrastructure you don't pick, layered under apps you do.
6. The value moves up the stack, again. Telecom revenue is a trillion dollars a year and the stocks have gone nowhere in 25 years, while the things built on top of mobile networks made trillion-dollar companies. The same pattern is likely for foundation models: huge revenue, thin margins, all the interesting wealth created by the people building on top.
7. When the product becomes commodity, distribution becomes the moat. Google and Meta are already spraying AI across every surface they own. OpenAI's "shipmas" sprint was an attempt to build a flywheel before that happened. Apple - whose 2024 vision of personal on-device AI was the most compelling demo of the year - is the last shoe to drop.
8. The anti-AI backlash is a fuzzy mess of real and unreal grievances. Some are true (electricity bills going up in specific places). Some are nonsense (data centers use 0.017% of US water). Some are an artist class watching the floor drop out on illustration commissions. Most discourse conflates them, which means none of them get addressed properly.
9. You can't predict which jobs are exposed. In 1997, the obvious safe job was taxi driver - what does the internet have to do with hailing a cab? Uber answered that. Today the supposedly safe jobs include personal trainer. Prop your phone on a rack, point the camera at yourself, ask AI to coach your form. Maybe that doesn't work. The point is you couldn't have predicted it.
10. Software engineers thought their job was the hardest to automate. It turned out to be the most transformed. Evans's read: engineers didn't realize that most of what they did was boring manual labor that could be automated. They thought it was creative work. The lesson generalizes - your job is probably not the thing you think it is.
11. The only useful response is to dive in. Going on Bluesky to shout about how evil AI is gives you a great feeling of moral superiority and accomplishes nothing. Walking into a law firm interview and saying "I think AI is bullshit and I'll never use it" is not the move. Submerge yourself in it. Come out the other side knowing what it can and cannot do. That's the only career insurance available.
12. AI corner - what Evans actually uses it for. Proofreading. Generating images for apartment redecorating ("here's a picture of this room, repaint it, add this rug, change the rug color"). Voice dictation that auto-transcribes to text. The general pattern: AI is good at the stuff computers used to be bad at, and bad at the stuff computers were good at. The boring precise retrieval tasks he most wants automated are still the things it does worst.
Amit B retweeted
On #WorldBicycleDay, riding my cycle in Vimannagar, Pune; these dangerous spots (potholes, hawkers, broken roads) show how PMC has completely ruined Puneโs cycle tracks.
Where are the safe dedicated lanes we were promised? This negligence is unacceptable @navalMH
Share your worst PMC ruined cycling spots in Pune ๐
@PMCPune @mohol_murlidhar
@VIRBIKE thanks for your cycle ๐
#PuneCycling #SaferRoads #Vimannagar #PMC_Pune #RideSafe #WorldBicycleDay #VIRBike





























