Shubh Karman Singh

🇮🇳 • swe • guitarist • posts about anything i find cool • iitian • i don’t vibe, i debug

[nisarga @ni5arga]

We managed to play the iconic Bad Apple video on CBSE's prod site!

[Shubh Karman Singh @qascade]

@ManikGarg826005 @ndtvindia Hahaha

[Shubh Karman Singh @qascade]

@DikshaKandpal8 Its the literal same system as cbse one, the password is also the same.

[Shubh Karman Singh @qascade]

@ajeetbharti @ni5arga Just click on the url once. You’ll see the magic 😭

[Shubh Karman Singh @qascade]

@ni5arga Like if you go in details, this is f’ed up at so many levels 🤦🏻

[Shubh Karman Singh @qascade]

@championswimmer Looks like you are blowing it out of proportion. He’s talking about research/academic things.

[Aran Komatsuzaki @arankomatsuzaki]

The non-English tax is real.

Sutton's Bitter Lesson, translated across languages and normalized to OpenAI English token count:

Hindi: OpenAI 1.37×, Anthropic 3.24×
Arabic: OpenAI 1.31×, Anthropic 2.86×
Chinese: OpenAI 1.15×, Anthropic 1.71×

Claude’s tokenizer charges a much higher linguistic tax.

[Shubh Karman Singh @qascade]

Never thought I’d see a day where people will leave github because of platform degradation

[Mitchell Hashimoto @mitchellh]

Ghostty is leaving GitHub. I'm GitHub user 1299, joined Feb 2008. I've visited GitHub almost every single day for over 18 years. It's never been a question for me where I'd put my projects: always GitHub. I'm super sad to say this, but its time to go. https://t.co/DQDemHdytV

[Todd of Mischief @AndToddsaid]

I wish they showed me stuff like this in school.

[True Indology @TrueIndology]

Gurudwara Bangla Sahib was originally Bangla (palace) of great Hindu King Raja Jai Singh.

When Aurangzeb summoned 6 year old Sikh Guru Har Krishan, Jai Singh protected the Guru & offered him residence here. He donated Bangla to Sikhs.

How Hindus protected & nurtured Sikhi🧵

[Shubh Karman Singh @qascade]

People are suddenly discovering security problems in MCP. I remember people pointing out on day one, when MCP went viral that its not designed with security in mind. Nobody bothered to care at that point, everybody was on rapid adoption spree.

[Cyber Security News @The_Cyber_News]

⚠️ Critical Anthropic’s MCP Vulnerability Enables Remote Code Execution Attacks

Source: https://t.co/7ghzzZJBKg

A critical flaw in Anthropic’s Model Context Protocol (MCP) exposes over 150 million downloads to potential compromise. The vulnerability could enable full system takeover across up to 200,000 servers.

Unlike a traditional coding bug, this vulnerability is architectural, meaning any developer building on Anthropic's MCP foundation unknowingly inherits the exposure from the ground up.

The flaw enables Arbitrary Command Execution (RCE) on any system running a vulnerable MCP implementation. Successful exploitation grants attackers direct access to sensitive user data, internal databases, API keys, and chat histories, effectively handing over complete control of the affected environment.

#cybersecuritynews

[Shubh Karman Singh @qascade]

There's a 10% chance that LLM can hallucinate results inferred from data. I hope people have factored that in.

[Aakash Gupta @aakashgupta]

Anthropic just turned the entire BI stack into a checkbox in a $20 subscription.

Tableau charges $75-$115 per user per month for what Cowork now does in a prompt. Looker's enterprise floor starts around $30K a year. Retool raised at a $3.2B valuation building exactly this: dashboards connected to your apps and files, refreshed with live data. They hit $120M ARR last October on it.

Claude Pro is $20/month. Cowork is included. The dashboard, the data connection, and the auto-refresh are now one prompt and a connector approval.

The pitch for $75/seat Tableau was always the connectors plus the visualization layer. Cowork has the connectors (Slack, Salesforce, Drive, Asana, Jira) and Claude writes the visualization in 30 seconds. The pitch for Retool was that engineers could ship internal tools in hours instead of weeks. Cowork ships them in minutes for people who can't write a SQL query.

Every internal tools team at every mid-sized company exists because "live pipeline metrics view" used to require a Retool license, a data engineer, and two sprints. That whole job description is starting to compress into a chat message.

Tableau, Looker, and Retool all priced on the assumption that dashboards are scarce and creators are rare. Cowork inverts both. Every employee can build their own dashboard. Fewer Creator licenses, fewer Explorer licenses, fewer Retool seats per company from here.

Cowork isn't the BI category killer yet. But it's coming for the seat count, and the seat count is the whole business.

[BuccoCapital Bloke @buccocapital]

“Sir, another 22 year old has found a job”

[Shubh Karman Singh @qascade]

I am trying to find the surprise here, when did a series of linear and non linear transformations on matrices, even remotely substantiated the idea of consciousness ?

[ℏεsam @Hesamation]

Google DeepMind researcher argues that LLMs can never be conscious, not in 10 years or 100 years.

"Expecting an algorithmic description to instantiate the quality it maps is like expecting the mathematical formula of gravity to physically exert weight."

[Shubh Karman Singh @qascade]

@aakashgupta > The companies who cannot afford to build now can

I agree, but most will have limited access to the best intelligence per token. Slowly, tokens are going to become a resource optimization problem.

[Shubh Karman Singh @qascade]

The more we try to move towards super intelligence, the more we see LLMs act like average pleb down the road. All the training data is generated by humans with bias. LLMs will always show human like behaviour because its just a next best “human text” predictor on steroids.

[Mo @atmoio]

AI is giving every CEO the same advice

[Hugging Models @HuggingModels]

Meet Ambuj-Tripathi-Indian-Legal-Llama-GGUF: a specialized AI model fine-tuned for Indian law. This isn't just another chatbot. It's a legal assistant trained to understand the nuances of Indian statutes, case law, and legal language. A game-changer for legal tech in India.

[Shubh Karman Singh @qascade]

@Priyansh_31Dec If your code base is a mess, agent is going to add to the mess. Every model has some sort of code style bias that may not match with your current code. Having agents.md instruction for that really helps. Use a separate agent for critiquing ideas and evaluating your ideas.

[Shubh Karman Singh @qascade]

@RaphaelDeLio Apparently, Claude works best with XML, Chatgpt works best with Markdown and Gemini works best with YAML