pratap

Engineer building smarter machines — from LLMs to trading bots.

[pratap @alphagradient]

Agreed.
I moved from @claudeai to @deepseek_ai v4 pro and flash combo recently for my startup.

[Chamath Palihapitiya @chamath]

Your margin is my opportunity: AI version…

The biggest surprise of 2026 is that the capability gap between the best open-weight/source models and the best closed models has narrowed much faster than the pricing gap. The pricing gap remains enormous while the capability gap is quite narrow.

What does this means in practice?

For a company consuming 1 billion input tokens and 1 billion output tokens per month:

GPT-5.5 Pro: ~$105,000
Claude Opus 4.8: ~$30,000
DeepSeek V4 Pro: ~$5,220
DeepSeek R1: ~$2,740

I asked ChatGPT what it thought about this and it answered as follows:

“If I were building a company today, the economic frontier would look roughly like:

DeepSeek V4 Pro / R1 for high-volume inference.

Claude Opus for premium agent workflows where reliability matters.

GPT-5.5 Pro only for workloads where its incremental capability demonstrably produces enough business value to justify a 20–40× token premium.”

Most CEOs have no idea that, instead of this nuanced approach, their teams are running amok internally by picking the most expensive models in most cases and burning through massive budgets with zero governance, audit ability and control.

As control planes like our Software Factory become more standard, you can expect the run rate revenue growth of the frontier labs to go down meaningfully and the revenues of the open models to skyrocket.

Why? Because we can implement the nuanced approach above and be agnostic to model - instead focusing on customer intent, model task and cost management among other things.

[Financelot @FinanceLancelot]

They changed the entire rules of the stock market specifically for Elon Musk's SpaceX IPO...

This is what they aren't telling you... and how it could cause the greatest collapse we've ever seen.

[pratap @alphagradient]

@NaveenGRao @unconvAI @databricks How dod you do it? Will look into this today.

[pratap @alphagradient]

@unusual_whales Foreign policy used to require at least a think tank. Now you just need a stonemason and a golf course pitch deck.

[pratap @alphagradient]

@elonmusk It also beats both of them at running out of range while towing. 130 miles with a full load and you’re a $100k stainless steel paperweight on the side of I-10.

[pratap @alphagradient]

@malinvested What conclusion @grok

[Jmail @jmailarchive]

We built Jikipedia, a new wiki that compiles Jmail data into exhaustive reports on key figures in the Epstein scandal.

It lists all recorded visits to Epstein's estates, each person's possible knowledge of Epstein's crimes and laws that they may have violated.

Here is Epstein's executive assistant, who sent >100k emails to him and was shielded from prison via a 2007 non-prosecution deal.

[Ro Khanna @RoKhanna]

This weekend the world's leaders will gather together for the Munich Security Conference.

I am calling on them to hold the Epstein Class accountable, including those in their own governments.

[Elon Musk @elonmusk]

@ianmiles There is much more to these files than has not been released

[Veer Syngh @Syngh_Unleashed]

Dear @ncert, this is nothing more than the usual Marathi rhetoric of equating tribute (tax) collection with political sovereignty. They falsely boast about ruling entire India even though levying tribute is not the same as actually ruling or administering a state.

Although I’ve discussed this multiple times in the past, allow me to summarise some of the key points 📌 :

🔸 The tax levied from kingdoms in the form of 𝘾𝙝𝙖𝙪𝙩𝙝 was more or less protection money or military extortion, compelling the local rulers (Mughal governors, Nawabs, Rajputs Rajas, etc) to pay it in order to prevent plundering raids by Marathas, the main victims of which were always innocent civilians.

🔸 This is not some random assumption but an academically attested view. To quote the words of reputed historians like Jadunath Sarkar :

“ 𝘾𝙝𝙖𝙪𝙩𝙝 𝙘𝙤𝙡𝙡𝙚𝙘𝙩𝙚𝙙 𝙗𝙮 𝙈𝙖𝙧𝙖𝙩𝙝𝙖𝙨 𝙬𝙖𝙨 𝙤𝙣𝙡𝙮 𝙖 𝙢𝙚𝙖𝙣𝙨 𝙤𝙛 𝙗𝙪𝙮𝙞𝙣𝙜 𝙤𝙛𝙛 𝙧𝙤𝙗𝙗𝙚𝙧𝙨 𝙖𝙣𝙙 𝙘𝙖𝙣𝙣𝙤𝙩 𝙗𝙚 𝙘𝙖𝙡𝙡𝙚𝙙 𝙚𝙫𝙚𝙣 𝙖 𝙨𝙥𝙝𝙚𝙧𝙚 𝙤𝙛 𝙞𝙣𝙛𝙡𝙪𝙚𝙣𝙘𝙚, 𝙡𝙚𝙩 𝙖𝙡𝙤𝙣𝙚 𝙖 𝙢𝙖𝙧𝙠 𝙤𝙛 𝙫𝙖𝙨𝙨𝙖𝙡𝙖𝙜𝙚. ”

🔸 There are many instances throughout Maratha history when their own chiefs, acting independently of any central authority (neither 𝙋𝙚𝙨𝙝𝙬𝙖 nor 𝘾𝙝𝙝𝙖𝙩𝙧𝙖𝙥𝙖𝙩𝙞), were compelled to pay tribute to the Mughal Emperor, the Nizam of Hyderabad & even the East India Company. So Marathas were ruled by Mughals, Nizam & British too?

🔸 Speaking of central authority, who exactly was the Emperor of this so-called ‘Maratha Empire’ ? Was it the Brahmin Peshwa of Poona or the titular Maratha Chhatrapati of Satara? Even the chiefs working under them — Holkar, Sindhia, Gaekwad, Panwar, Bhonsle — fought against each other like bitter enemies throughout their history.

🔸 Even the 𝘾𝙝𝙖𝙪𝙩𝙝 demanded from Rajputana during the regency of Mahadji Sindhia (𝙑𝙖𝙠𝙞𝙡-𝙞-𝙈𝙪𝙩𝙡𝙖𝙦 of Mughal Empire) was legitimised in the name of the Mughal Emperor only. Rajput states had been holding back tribute to the Mughal government for years, while Marathas came demanding for it on behalf of their masters.

🔸 How many North Indian rulers in general (or Rajput Rajas in particular) formally attended any Maratha 𝙙𝙖𝙧𝙗𝙖𝙖𝙧 (court) even once, let alone on a frequent or regular basis like subjects are ought to? Did any of them ever render military service under Marathas or assigned an official rank in the Maratha bureaucracy (assuming it was one)?

🔸 No, not at all! It was the exact other way around with Maratha leaders often visiting the Mughal court (or sending their direct representatives) to appeal for 𝙈𝙖𝙣𝙨𝙖𝙗𝙙𝙖𝙧𝙞 or 𝙎𝙪𝙗𝙚𝙙𝙖𝙧𝙞, addressing themselves as loyal slaves of the Mughal Emperor in their royal seals & even going as far as minting coins in his name.

🔸 Marathas did not even care if any other power attacked & looted their so-called vassals as long as they were getting their own share of ‘𝙝𝙖𝙛𝙩𝙖-𝙫𝙖𝙨𝙤𝙤𝙡𝙞’. If we are to go by their logic, then underworld criminals & political goons — like Dawood Ibrahim or The Thackerays — are also ruling sovereigns of Maharashtra.

[pratap @alphagradient]

These are 10,000-year-old voices from our past, slowly fading in the open.

Bhimbetka deserves world-class preservation, not quiet decay.

Heritage once lost is lost forever.

@PMOIndia @MinOfCultureGoI @BJP4MP

[pratap @alphagradient]

Companies have started to price out consumers from building their workstations running open source language models. SSDs, RAMs, GPUs should be bought as much as you can while they are still available. #AI

[Ahsan @AhsanKharbai]

What a time to be Alive..

You can get an SSD at 16 Lakhs or an SUV at less than 16 Lakhs

Which one are you picking?

[DFarmer @OGDfarmer]

Just give us back Twitter man, this shit’s unbearable.

[BythePeople @BythePeople18]

Stephen Miller is of Ashkenazi Jewish descent. His great-grandparents escaped Russian pogroms and the Holocaust. He is from a family of immigrants! I, being Jewish as he is, cannot comprehend how he can be so hateful and anti-immigrant.

[The Alex Nowrasteh @AlexNowrasteh]

Good for him. Glad to see a conservative with guts stand up to all the anonymous race-obsessed troll accounts.

[pratap @alphagradient]

@virsanghvi Which minister has 100 IQ? Or you were pointing to combined IQ of Delhi BJP govt?

[Jay Yang @Jayyanginspires]

You can just do things

[pratap @alphagradient]

Real work is stateful and full of hidden constraints. It shifts constantly. It’s non-i.i.d. Benchmarks aren’t.

[VraserX e/acc @VraserX]

A former Google DeepMind research scientist is basically saying out loud what many engineers feel right now. Gemini 3 is a solid web dev helper, but when it comes to real coding it often feels overfit for benchmarks instead of real world work.
Great at demos, shaky in the trenches.

[pratap @alphagradient]

Multi-head attention isn’t parallel computation.
It’s parallel nonlinear decision boundaries. #attention

[pratap @alphagradient]

Hot take: most explanations of multi-head attention overcomplicate it.

You can replace almost all of them with one fact:
QKᵀ (full) = Σᵢ(QᵢKᵢᵀ), but
softmax(QKᵀ) ≠ Σ softmax(QᵢKᵢᵀ).
That “≠” is why multi-heads exist.
Each head normalizes its own similarity matrix in private.
Neural networks do NOT like shared custody.