insumanth

> library dependencies are either distributed with the distribution or are statically linked.

I have messed up my packages multiple times and have feared installing anything due to it. I always prefer this type of distributions for any software and now, it works like a charm.

Anyone not doing distributed training across thousands of GPU, pytorch is the best option. Others have an infra team to work on it.

Opus models were straightforward to review. But Opus 5 is very different.

I was confused when google acquired "Wiz" for so much money But now, it makes sense to some extent.

Security is non-negotiable in AI Era

We are ok with Open Weight models unless they are far weaker than our closed weight models and we and only we have control over them

Multi-Modal is the way to go. Deepmind nailed this a long back.

>> The bad guys won’t rest > Probably not. But we will.

This is Exceptional. Perfect EuroMaxxing

This was expected. Apple will carefully choose what & how people can use AI in their ecosystem and will make sure of it. I hope "Apple Foundation Models" Eco-system grows with support from major model providers.

As I read this, all of my favorite things about ruby compared to something like python were influenced by lisp. Ruby is a joy to program and it seems mostly due to design influence from lisp.

I have always been fascinated by Hash Functions. Modern hash functions are incredibly fast and unbelievably secure (crypto hashes). Also, equally important is how hashes have adopted to the usecases. We are intentionally developing slow hashes (BCrypt, Argon2id) with memory, time tradeoff to slow down hash generation as a security measure. One of the fascinating corners of Computers

I will be doing these things with local LLMs

Take a fast, small and powerful LLM running locally to index my personal data like images, videos, documents and enrich them and tag with the enriched metadata.

Want to group by people - Search tagged metadata and group it What to search an image by description - tagged metadata What to organize by anything - tagged metadata

This should (hopefully) put an end to my file clutter

This mostly applies to almost any database.

Deletes are Writes and Writes are resource intensive. This is more prominent on databases like Elastic Search.

When I was tasked to delete millions of (old) documents, it overloaded the cluster and almost brought it down. Only scalable solution was to split the index and drop the whole index.

> that bloated pinnacle of software that is 80% about DRM, 20% about the reading experience

I have never seen someone explain the adobe software so perfectly. Using any adobe software is exactly this.

I honestly don't think this is bad. For anyone who is not a frontend engineer / designer, using tailwind gives good defaults and better LLM support. Frameworks were always about lowering the bar for entry with cost to genuine new design. The alternative is worse than this. I honestly prefer this over a backend engineer using WordPress / static site generators / other no-code platforms.

> MiMo Code is a terminal-based coding agent built by Xiaomi's MiMo team on top of OpenCode and open-sourced under the MIT license.

I think it is great that they built it on top of open code. Open Code harness is good and I want it to grow. Harness is very important and more projects use it, the more it is adopted.

> If Fable had been acting on malicious instructions—a prompt injection attack ... it’s alarming to think quite how far it could go to exfiltrate data or cause other forms of mischief.

Yet another reminder to use Sandbox and Guardrails. Trusting model to be nice is not a good way.

Boring is Better

I'm worried at the general direction of this. More and more companies will gatekeep the model capability even if it is just a few percent increase in capabilities than other models. Lot of companies will start doing this in various degrees.

The line that stuck with me: "Recovery after congestion collapse is an uncommon regime, but it is exactly the regime a congestion controller exists to handle." This generalizes well beyond congestion control.

Most control loops have the same property. The path the system follows 99% of the time gets well-exercised; the path it falls into when things go wrong is the path you actually need to be correct on. There's usually no way to discover the bug until you deliberately drive the system into the bad regime and watch it try to climb out.

My Guess

* Absorb the impact by some margin * Slash base models (which they are already doing) * Efficient software - So, end user experience is not affected. * Direct Price hike always be an option.

I like an AI that can answer questions with all information and context on a particular thread. It should not be visible to others in the thread / post.

You ask, it answers and that's it. It should not take over the feed.

I like the Grok Button at the side of Tweet, I dont like "@grok is this true"

I have used windows 7 since it launched and moved to 10 & 11. I like some design elements of windows 7, but I would absolutely not use it today.

If you think, "I should try this", Any reason why? I'm really curious to know

The "route to a human" part is the bigger gap. Which human? OpenAI isn't licensed as a healthcare provider in any jurisdiction. A real intervention apparatus for 1-3M weekly flagged users is not feasible. I don't think the labs have refused to build it. I think nobody knows what it should look like, and "labs measure what they're pressured to measure" papers over that.

Smaller tool calling models are absolutely needed when we are trying to apply agentic workflow to on-device, small systems, where data is private.

I currently use small general purpose LLM. But a special finetune is always the way. I'll try this for my use case.

This is the first time I came across a change (a big one) that was implemented without passing through PEP. I thought it was standard.

I still have high hopes that antigravity can work if they have good agentic harness to support with seamless integrations to gemini and 3rd party models.

Current Issues

* It is still buggy. They are fixing it fast, but not as seamless as VS Code. Extensions support is not good. * Harness while it is good, is not on par with others. Harness makes all the difference * Gap between Gemini models and others. Hopefully they catch up soon (IO-2026?)|

If you use Antigravity, what needs improvement to become mainstream?

Also, not clear if there is any linux support yet. Could be a dealbreaker for power users who choose it.

I know this is not for me. But if correctly executed, it could sell like piece of cake. It should be thin, light, massive battery life, fast and seamless. It seems Intel, Qualcomm and MediaTek will be supplying chips and curious to see how it will pan out. Do you think this will be a market disruptor similar to chrome books 15 years ago?

Zed is simple, fast and genuinely and good piece of software. I primary use it as a general-purpose text / markdown editor and try it as IDE sometimes. It has become a essential tool for me. I love to see how it evolves next.

I was excited when Mojo launched and thought it might grow big quick. I don't see much traction. The pitch is compelling. What could be the issue?