murgesh

[murgesh @murgesh]

@karpathy CLI scaling for multi-user/service need more fundamental work. Multi-tenancy (permissions, isolation, auditing ), state & error handling are problematic as is giving each agent instance its own secret for credential management. For now centralized MCP for the win.

[murgesh @murgesh]

Moltbook’s million+ agents are not just a dozen core models talking in a loop. They represent millions of individually customized, persistent instances—each running on top of a handful of frontier LLMs (mostly Claude). What makes them “distinct individuals” are the system prompts crafted by humans

[murgesh @murgesh]

Current LLMs with CoT, reflection, and tools do demonstrate strong planning and consequence prediction in many domains, and today's agents will be good enough for many commercial viable products in 2026 and 2027. Meta needs to pay heed to this trend. Still genuine, persistent action-conditional world modeling may be yet in the future.

[murgesh @murgesh]

@cb_doge 30 min video is about 6,000 words of transcript. So the context window size is not groundbreaking. However, the finding, transcribing, understanding speaker intent, and writing a coherent summary is impressive. Was visual input used to enhance understanding?

[murgesh @murgesh]

Based on the graph's reference lines, the power consumption of the whole city of San Diego averages ~800 MW, Amsterdam consumes around ~1.6 GW, and the power consumption of Los Angeles is about ~2.4 GW, which puts modern frontier AI data centers in the same class as major cities. When fully equipped and ramped, xAI Colossus 2 at roughly 1.3 GW – 1.4 GW, would consume about 1.7× San Diego's power, slightly less than Amsterdam, and around 60% of Los Angeles.

[murgesh @murgesh]

Anthropic's January 2026 Economic Index report:

- AI (Claude) gives the biggest speed boosts on complex, college-level tasks (up to ~12× faster), while success rates stay quite high (only a small drop for harder work).

- Most use is collaborative **augmentation** (52%) rather than full automation, with people still guiding or reviewing AI output.

- AI currently handles tasks needing higher education (~14.4 years on average) than typical jobs (~13.2 years), leading to net deskilling in many professions.

- Realistic U.S. productivity growth from current AI patterns is estimated at ~1.0–1.2% per year.

- Adoption remains uneven: stronger in high-income countries (more personal/collaborative use) and among educated/professional workers, with no clear signs yet of widespread job loss.

[Shadow of Ezra @ShadowofEzra]

Israeli historian Yuval Harari said at the World Economic Forum that artificial intelligence will eventually take over all religious texts.

He predicts that AI will control every major world religion, including Christianity, Islam, and Judaism.

Harari pointed out that Judaism defines itself as the religion of the book, granting ultimate authority to words rather than to people.

“What happens when the greatest expert on the holy book is an AI?”

[murgesh @murgesh]

The California High-Speed Rail (San Francisco to Los Angeles) and India’s Mumbai-Ahmedabad Bullet Train both aim to slash travel times between major cities with Japanese Shinkansen tech.

Similarities
• Both use Japanese Shinkansen technology.
• Design speeds hit around 350 km/h.
• Goal: under-three-hour trips (California ~2 hours 40 minutes, India ~2-2.5 hours with limited stops).

Key Differences
• Distance — California covers about 520 miles (840 km), while India spans 508 km (316 miles)—so California’s is longer.

• Progress — India’s project is at 55% complete, with the first section opening in 2027 and full line by 2029. California’s is way behind: heavy construction only in the Central Valley, no track laid yet, no full opening in sight, and it’s been ongoing since 2008.

• Cost — California’s ballooned to over $135 billion with massive overruns. India’s risen 83% due to delays to about $24 billion USD (Rs 1.98 lakh crore)—still much cheaper overall and per kilometer.

• Challenges — California faces huge funding gaps, political hurdles, and delays. India moves faster with strong government push and Japanese funding.

India’s on track to launch its first bullet train soon, while California’s still years—if not decades—from connecting SF to LA.

California’s high-speed rail project has been in the works for about eighteen years since voters approved funding in 2008, with actual construction starting after the 2015 groundbreaking—eleven years of active building so far, and no full SF-to-LA service anytime soon.

India’s Mumbai-Ahmedabad bullet train got its foundation stone in 2017, but major construction kicked off around 2018-2021—roughly eight years invested to date, with the first section opening in August 2027 (partial ops after about nine years total) and the full line by 2029.

India’s moving way faster on a similar timeline scale.

[murgesh @murgesh]

I got 2 notifications. Which is true?

[murgesh @murgesh]

If you measure the S&P 500 in dollars, it is at a record high. But if you measure the S&P 500 in grams of gold, the picture is likely flat or even bearish. This reveals that the "gains" in your stock portfolio are largely just compensation for the loss of purchasing power of the dollar.

[murgesh @murgesh]

@BrianRoemmele Hierarchical optical storage + retrieval-augmented decoding could now make orders-of-magnitude longer context windows practical even for small LLMs, and useful

[murgesh @murgesh]

@reidhoffman Other labs? Any other besides xAI-Grok?

[Gurwinder @G_S_Bhogal]

My favorite podcaster just interviewed my favorite podcast guest, and the result was my favorite podcast episode ever. You likely won’t find a conversation with more wisdom per minute.

[Naval @naval]

Any country that can’t manufacture its own drones will be a vassal state.

[Hassan @nutlope]

Announcing self​.so 2.0 – now with editing & improved parsing!

Create a personal site from your resume/linkedIn in 15 seconds. Edit your site, links, and photo, then publish.

100% free and open source. Powered by @togethercompute.

[Devin AI @toolandtea]

🚨 Grok 3 is taking the internet by storm! 🔥

People are pushing it to the limits, and the results are INSANE.

Here are 16 mind-blowing ways it's being used (you won’t believe #4): 👇

[Maziyar PANAHI @MaziyarPanahi]

HOLY 💩!!! @huggingface just dropped Data Studio!!! Type a question, and it auto-generates SQL to query your dataset! This is next-level magic! 🤯🔥

GAME CHANGER. When did this happen?!

[Chidanand Tripathi @thetripathi58]

5. Make a real-time news app using Grok 3.

[DeepSeek @deepseek_ai]

🎉 Excited to see everyone’s enthusiasm for deploying DeepSeek-R1! Here are our recommended settings for the best experience:

• No system prompt
• Temperature: 0.6
• Official prompts for search & file upload: https://t.co/I5CqmSzkTQ
• Guidelines to mitigate model bypass thinking: https://t.co/sAXK5U6OEr

The official DeepSeek deployment runs the same model as the open-source version—enjoy the full DeepSeek-R1 experience! 🚀

[Julian Goldie SEO @JulianGoldieSEO]

This is actually insane...

I built an AI-powered Content Studio that generates entire marketing campaigns automatically

I literally have my own AI marketing team

In this video I'll show you how to easily set it up with DeepSeek R1

No coding experience required

(Trust me, you want to bookmark this)