←
Siva4,159 posts
Siva avatar

Siva

@ergodicthought

“the best prize life has to offer is a chance to work hard at work worth doing” | “... specialization is for insects”

◷ Joined October 2008▧ 44 Photos and videos
Mike Rea avatar
⟳ Siva retweeted
Mike Rea@ThenIf_·Nov 16

"Lilly made 64% of its 2025 sales from medicines approved in the last five years. Sanofi is second, at 14%. Amgen is third, at 12%. Everyone else in the top 20 is in single figures, and Gilead and Novo Nordisk are at zero. The revenue age tells the same story, even if it does it more quietly. Lilly’s is 6.6 years. For most of the group, it sits between eleven and thirteen." @DaveARicks The Aftermath Index asymmetriclearning.substack.…

Post attachmentPost attachment
Siva avatar
Siva@ergodicthought·10h

Great to have some local competition 🙌

We launched Jarvislabs Token Factory two months ago, running on GPUs in India, and have been heads down improving it ever since. We benchmarked the same three open-weight models on Sarvam's Token Factory and ours. We're on par with or faster than Sarvam across speed and latency metrics, while offering lower pricing. - GLM 5.3 Flash is almost 2x faster. - Gemma 4 31B and DeepSeek V4.1 Flash are slightly faster, with significantly lower pricing for input, cached, and output tokens. And we're running on H200s, not B200s btw.

Post attachment
Siva avatar
Siva@ergodicthought·12h

Very clever. Another manifestation of how you use a "world model" to attribute outcomes to specific control inputs, and solve counterfactual reasoning.

We might’ve killed retraining for harness changes. Our model broke on a new harness. We didn’t retrain it. We edited 5,064 weights. Out of 4 billion. 20.1% -> 88.2% No gradient descent. Old-harness accuracy didn’t move. DAWE(Data-Attributed Weight-Editing) is fully open 🔥

1:10
Amar Govindarajan avatar
⟳ Siva retweeted
Amar Govindarajan@amargov·Nov 16

For years, Indians have paid well above world prices for solar panels. A 40% customs duty on imported panels, 25% on cells, and rules letting only Indian-made panels into government-backed projects kept cheap Chinese imports out. This is sound industrial policy. CAVEAT: The idea was that Indian makers would use the profits from the import protection to climb up the value chain. You know, learn to not import tech. Their own annual reports show how that went. What it earned the companies: - Five of India's biggest listed solar makers -~₹7,300 cr combined profit in FY26 - The largest of them: over ₹10 cr profit a day - Another kept ₹19 of every ₹100 of sales as profit - The market leader's own annual report credits import duties with helping Indian makers "consolidate their market position" What they spent on learning (FY26 annual reports): - All five together -~₹125 cr on R&D - under 2% of their profit, or about six days of it - The biggest spender among them: about 50 paise per ₹100 of sales - The market leader: 18 R&D staff in a workforce of 10,000+ - One printed no R&D figure at all, in a year it made ₹1,082 cr profit - Another's parent company reports R&D: "Nil" - For scale: -First Solar (US) spends ~₹4.50 per ₹100; -LONGi and JA Solar (China) ~₹6 - while making heavy losses What that leaves India with: - Wafers for only about 1 in every 100 panels it makes - The market leader's latest technology import: a 10 GW ready-made cell line (2025) - It is "sending our engineers to China for structured training" - 16 October: China proposes curbs on exporting 12 solar technologies - The industry's answer: shop for machines in Germany, Japan and Korea instead Consumers paid the premium. The companies banked the profits. Research got less than 2% of them. (There is also a identifiable way out - how Indian solar tech players are struggling for orders - etc - this is in the article attached in next link)

Post attachment
Siva avatar
Siva@ergodicthought·Oct 31

Translation: Now that everyone is used to not seeing the reasoning tokens they're paying for, OpenAI is well-positioned to basically fake the amount of thinking to justify value perception and market segmentation. OpenAI wants to pivot from a commodity services business model (token pricing) to value based pricing, and Sam is trying to sell us this new reality.

Sam Altman says tokens are "fake from of a business from value point of view." Open ai almost stopped charging on tokens. Sam Altman: "We almost made what would've been a bad call, to say we're not going to price our next model in tokens at all. You're just going to give it a prompt, and we're going to give you a price back. It's going to be like AWS to you. Just do something and we send you a bill." "But the customer response on that is so negative. They're like, 'I need to know what I'm paying for somehow." "I asked a few people, 'Would you rather pay more for the same output, but have it priced in some number of consistent tokens ?' Even if we inflated the tokens, they're like, 'Oh, I'd much rather you inflate the tokens and pay more, because I know there's a real thing there.' "People don't just want us to give them a price, because they feel like over the long term they're going to get screwed somehow." "I would rather not make up a fake unit. Tokens are a kind of fake unit. I mean, they're real in some sense, but they're not a good unit." "They're fake from a business value point of view."

2:22
Siva avatar
Siva@ergodicthought·Oct 30

🤔 Is this the backstory of claudeslop? Every 50 tokens fighting to be similarly load-bearing without long form coherence.

How to get ppl to read your stuff: People ask to review their blog posts and my response is usually some polite form of "it doesn't matter, no one will read it." It's a rough thing, but it's true. John Grisham in a 2017 interview said he's irritated by new authors debuting 1000 page novels because they haven't earned the right to publish books that long. That was 2017 and this is now and you haven't earned so much as a right to publish a screen-long blog post Scott Alexander can do it because he wrote interesting stuff for many years. Pg can do it because he wrote useful stuff. Dario Amodei can do it because he built probably the most important technology in history of the world. But you didn't do any of these things (btw be honest did *you* read every word in machines of loving grace from start to finish?) If you're new you have a baseline right to 15 tokens of ppl's attention. If these are really good tokens, you've earned the right to another 50. And if those are really *really* good, you've got another 50, and so on. See, every 50 tokens an interrupt fires in ppl's brains and they decide from scratch whether to keep reading or switch to literally anything else on the internet So don't worry about spelling, grammar, structure, thesis, capitalization, punctuation, or any of that writing stuff because it doesn't matter. You must produce valuable tokens, 15 first, then 50 tokens at a time. Self-review for that, you'll know in your gut if you're doing it. Attention is all you need Btw openai tokenizer tells me this post is 9 -> 95 -> 76 -> 84 -> 65 -> 42. So i'm breaking my own rules a bit, i need to get better

Arnav Bansal avatar
⟳ Siva retweeted
Arnav Bansal@itsarnavb·Oct 29

Launching Junction: come speed-run the semiconductor tech tree this December in Bangalore. We’re flying in students from all over India 🇮🇳 to a 2-week fabrication retreat at @_lagrangepoint to build equipment, design chips and engineer processes, with our friends at @hackerfabindia. India still needs thousands of semicon engineers -- the ecosystem to produce them isn’t mature enough yet -- we're changing that. Current applicants have built vacuum chambers to study plasma physics, RISC-V processors from scratch, scanning tunneling microscopes, et cetera At junction you will: 1. Propose a project 2. Figure out a plan with the Junction team. 3. Iterate fast towards a working prototype, with help from the Junction team 4. Be mentored by top semiconductor hackers and specialists. 5. Demo to industry leaders, founders, investors, and your fellow craftspeople Flights, accommodation and project expenses? All covered. Apply on the link below

2:40
Siva avatar
Siva@ergodicthought·Oct 29

When engineering any complex system, evals are the lens that shape the builder's perception -- and therefore an upper bound on how much one can reliably improve that system. There is a reason that scientific breakthroughs usually follow new measurement techniques. Its role in the dev process is more like homework assignments and less like the final exam. To think of testing as the last step, rather than a close associate throughout the development process, is to fly blind during development -- instead of with regular and high-fidelity feedback. PS: Observability is what makes certain systems (eg. physics) easier than others (eg. biology), and why the traditional "engineering" fields are as reliable as they are.

Why India's Deep-Tech Needs Engineers Who Can Invent New Tests Karan Kamble writes exhaustively about the need for a culture and ecosystem of owning testing paradigm in deep tech. I’ve shared my experiences in US and India and my perspectives on the topic. Thanks for the opportunity @astrokaran swarajyamag.com/technology/w…

Siva avatar
Siva@ergodicthought·Oct 28

Reminiscent of the Tamil saying "leaving the elephant's trunk and catching its tail" Better late than never, though. Good to see Indian IT waking up to the moment. They might just manage to do the right thing, after having tried everything else.

After HCLTech led Sarvam’s $300m round in June at a $1.5b valuation, Cognizant is reportedly in talks to join the next one, which would value Sarvam at $3.5 to 4 billion, more than double, with large Indian enterprises lined up to anchor it. For an industry trading near five-year lows on fears that AI agents will automate away coding, maintenance and back-office work, good to see them backing homegrown model builder is both a hedge and a bet on the future. via @EconomicTimes

Post attachment
Siva avatar
Siva@ergodicthought·Oct 26

I'm using this (the paid ZDR version) for a slice of my token needs and it's been very pleasant so far!

We're really, really excited: we're now serving DeepSeek V4.1 Flash at Gödel Machines, hosted in India. 1M-token context, tool calling, reasoning, and it's still free to use. That does mean we can no longer run Qwen 3.8 27B. We don't think you'll miss it. DeepSeek is something else, and we can't wait to see what you build with it.

Siva avatar
Siva@ergodicthought·Oct 24

It's interesting the kind of corpses that lie buried in ML algorithms, because the field incentivises scaling up more than getting the basics right.

Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:

0:07
Siva avatar
Siva@ergodicthought·Oct 24

And it shouldn't! Society only functions if actors have skin in the game for their actions; can't just socialize away the harms.

dan@irl_danBOct 23

my only comment is “we warned you that the thing we’re building would hurt you” will not fly with the general public

pH avatar
⟳ Siva retweeted
pH@pHequals7·Oct 19

happy AIyudha pooja to those who celebrate

Post attachment
Prathyush avatar
⟳ Siva retweeted
Prathyush@prathyvsh·Oct 18

Meanwhile Bret Victor on mechanistically generated interfaces — “Is this really the civilization we want?”

Post attachment

I'm still waiting for a Bret Victor level demo of new interaction paradigms enabled by LLMs