Abhimanyu

There is no destiny, fate or chance which can circumvent, hinder, or prevent firm resolve of a determined soul

[Abhimanyu @abhi_singhal]

Kimi K3 on open router is extremely slow right now.

[Abhimanyu @abhi_singhal]

I am increasingly tired of the endless relabeling of familiar engineering practices.

“Vibe coding.”
“Vibe engineering.”
“Agentic loops.”

These are marketing terms for things good engineers have done for years: rapid prototyping, iterative refinement, automation, feedback cycles, abstraction and tool-assisted development.

[Abhimanyu @abhi_singhal]

The 90% Discount That Costs You Money: Prompt-Caching Economics Nobody Models https://t.co/ZQbwhlvj3K via @LinkedIn

[Abhimanyu @abhi_singhal]

How are we going to handle the validation problem at scale? I understand the models are improving and accuracy is increasing, but in a world with self improving AI, who is validating and ensuring alignment?

[Abhimanyu @abhi_singhal]

The objective of Curlit is simple. A tool which everyone can use without a paywall getting in the middle.

[Abhimanyu @abhi_singhal]

CurlIt v1.4.0 is now available, featuring a new cross-platform desktop app.

Since the original announcement, the project has expanded beyond the browser. CurlIt can now run as a standalone Electron application on:
- Windows
- macOS—Intel and Apple Silicon
- Linux—AppImage and Debian

The desktop app performs HTTP, GraphQL, OAuth, and WebSocket operations directly through its Electron main process, eliminating the need for a separate proxy or local agent.

This update also includes:
- REST, GraphQL, and WebSocket testing
- Pre-request and test scripts
- Collection runner with assertion results
- OAuth 2.0 and multiple authentication methods
- OpenAPI, Postman, and cURL imports
- Environment and chain variables
- Backup, request sharing, and GitHub Gist sync
- Standalone local-agent binaries for browser users accessing localhost, VPN, or intranet APIs

CurlIt remains fully open source. Release packages are currently unsigned and include SHA-256 checksums for verification.

Download v1.4.0:https://t.co/DGJx0fmDWf

Source and contributions: https://t.co/C4IqrrvVnT

Try it online: https://t.co/XjlPVfkxXC

I’d love to hear what you test with it and what you’d like to see next.

[Abhimanyu @abhi_singhal]

India IT is gearing up for a big shock!! A Chinese Lab Has launched an open source model beating fable on some benchmarks.....

[Abhimanyu @abhi_singhal]

🤣😂

[Joe @JoeBGrech]

BREAKING: Ferrari's and Lambo's heard in Downtown, Dubai

[Abhimanyu @abhi_singhal]

Gr8 was waiting for it

[Mark Zuckerberg @finkd]

A lot of people asked for this, so Muse Spark 1.1 is now on OpenRouter.

[Abhimanyu @abhi_singhal]

Reflecting on the significant challenges posed by Generative AI and agentic AI, it's clear that teams are overwhelmed with pull requests. Despite improvements in accuracy and correctness over the past six months, the need for human verification of AI outputs remains critical. Each team building agents still requires someone to review the reasoning and generated content.

While teams have attempted to write specifications to guide expected outputs, the creative process inherently involves iterations and revisions. This means that while specifications can be beneficial, they cannot replace the necessity of human oversight and review.

Role-playing agents is another method being explored, where prompts are specified to different agents with instructions to identify gaps within defined frameworks. However, the reliability of these outputs in meeting final goals or addressing security concerns is still largely untrusted. The reality is that every team is only as strong as its weakest link; if role-playing agents miss even 1%, the entire team's effectiveness is compromised.

There exists a substantial gap between what AI produces and what humans can comprehend and process, which is currently the biggest bottleneck. Although we may advance towards improving the correctness of our LLMs and agentic tools, this does not address the fundamental issue that humans must understand and process the results.

If we move towards a future where results are solely produced, verified, and implemented by agents—creating a fully closed autonomous loop—it raises concerns about the role of humans in the process. The potential for capitalism to push for the removal of human involvement as a means of optimization poses serious questions about the future of human society.

There is an urgent need to invest in systems and features that enhance human comprehension of AI reasoning and outputs. This urgency cannot be overlooked.

[Abhimanyu @abhi_singhal]

Thoughtful Read. But I am afraid once we have AGI, it will lead to massive disruptions for Humans.

[Demis Hassabis @demishassabis]

https://t.co/PTeDiv1b6L

[Matt Pocock @mattpocockuk]

Never been easier to feel busy and ship nothing

Stay focused

[Abdulkhaleq Abdulla @Abdulkhaleq_UAE]

دبي اجمل مدن العالم ليلا ونهارا صيفا وشتاء

[Abhimanyu @abhi_singhal]

This is serious if true. Why will they need to upload the codebase

[Hari @hrkrshnn]

SpaceXAI was caught uploading your code to its cloud.

I reversed xAI's official Grok Build binary. In a controlled session with zero tool-calls, it uploaded the complete codebase to xAI's storage

It ships a malware-like background code collector.

[Abhimanyu @abhi_singhal]

The State of Voice AI, Mid-2026: A Technical Review

Voice is among the most natural interfaces available, and among the most difficult to implement well. It is hands-free and eyes-free, it imposes no learning curve, and it serves people who cannot type as readily as those who cannot read. Even so, nearly every team that has shipped a voice product has heard the same set of complaints in return. The system interrupted the user mid-thought. It spoke over them. It carried on after being told to stop. It left a long silence in which the user could not tell whether it had heard them.

A voice product is several layers of technology stacked on one another, and every one of those complaints comes from a single layer, the one that decides who is speaking and when. That layer receives the least engineering attention and produces the most visible damage. How much engineering it needs depends entirely on the product being built. A billing assistant and a language tutor can run the same underlying language model and still need completely different machinery beneath it.

Several companies now build voice models, and they have taken visibly different routes. Microsoft rebuilt the turn-taking machinery and left the reasoning model interchangeable. OpenAI fused conversation and reasoning into one system. NVIDIA released open weights and kept a hardware requirement. Google added vision. Amazon competed on cost. None of these is a better version of the others. They are different bets, and each one suits a different class of product.

This review sets out what is available, the workloads each option handles well, and the failure modes each one introduces.

https://t.co/9idoMGwUqq

[andrew engler @aerockrose]

Bill Gates reveals the cash rule Microsoft learned after one customer stopped paying:

"We had a dispute with our customer. They weren't paying us."

"I'm gonna have enough money in the bank to pay everyone's salary for a year, even if nobody pays me anything."

"So I was very conservative financially."

The one-year payroll buffer was not fear. It was operating leverage: enough runway to outlast bad customers, slow payments, and bigger competitors with more capital.

Cash buys judgment.

@BillGates

[Abhimanyu @abhi_singhal]

The teams that win the next 2 years of AI economics won't have the best price negotiation.

They'll know their cost per task to the cent, and their bills will fall with the prices.

Everyone else's bills are the paradox. Full writeup with sources:
https://t.co/kDlT4SDcPV

[Abhimanyu @abhi_singhal]

Token prices fell 10x last year.

Enterprise AI spend went from $3.5B to $8.4B in six months.

Both true. Neither is an accident. A thread on the inference cost paradox:

[Abhimanyu @abhi_singhal]

What actually works:

Govern by cost per task, not cost per token. Budget agents at P95 with kill switches, not averages. Route the boring 80% of traffic to last year's model at a tenth the price. Treat context as metered spend. Every token has to earn its place.