mrgaro

I think you could do that already. Just store your weights in a table...

Luckily this depends on manufacturer. I've been buying La Sportiva BoulderX approach shoes for around 10 years. I have like my 7th pair or something.

I measured this on my past claude code sessions. I have a dataset with full LLM requests and responses and I ended up to an estimated 5-7% savings on tokens. Also with my primarily Opus 4.8 usage, only 5.9% of unique reads even qualify for the bulk-read operation and 58% of the reads were already targeted.

Any DGX Spark users in this thread? What's your favourite model to run on it?

Because everybody has the same system prompt, the KV caching will make this a non-issue. The only cost is the reduced max context length.

How do you like the AMOLED screen? I have a Fenix 6, which has the old screen. While it doesn't have a flashlight, I'm very happy with the ability to toggle the backlight on with a single button, which is more than enough for me to see in the dark where I'm going.

I feel that I'm lucky as my Fenix 6 still works as I would not want to switch to an AMOLED screen, which draws too much power to be kept on 24/7.

How would you compare it to Opus?

That's so cool. I also thought a bit what it would mean to monetise the checklist/packing list app, but I just don't see much business case with it, with all the QA and support requirements.

Curious, what have you built? I just wibe coded a packing/checklist app, a llm-powered meal planner and a very specialised app for tracking LAN party network building.

Have we reached the capability of a local STT+LLM system being constantly listening for normal speech in a room and being able to understand when the human is addressing the system instead of talking to another human?

This! When you have the pictures on your disk just use immich-go to import them.

Great article! After reading I started to think there must be some purchable items which would demonstrate colours outside P3 colorspace. It would be cool to hold one on your hand and experience how a photo of it just cannot do justice.

Anybody know any links to webshops for such items?

In order to get enterprise agreement you need to pay per token for Claude.

Another examole which is trivial with MCP but hard with cli binaries: blocking certain commands, such as write operations from the agent. With MCP your client can easily have a blocklist for commands, but with cli you would need to code custom logic for each cli separately.

MCP has a great advantage over agent using cli: MCP is much easier to secure so that it's hardwired that the agent can only call the pre-configured MCP server. We run our agents so that they don't have access to public internet, so they could not run any cli commands. It's all either built-in agent tools, or 3rd party mcp servers. The agents never have access to any credentials, which makes them much more safe to use than a claude code running in yolo mode with fetching random cli binaries from the web.

I'd love to get this as a self-hosted web service, so that I could access my infinite terminal canvas from any browser. It should of course held the sessions in the background even when I'm not connected.

Often, especially on competitive games, the server is basically a full client, but just without graphics. The server will often run physics simulations etc, so that it can validate that nobody is cheating.

Sure, in some cases you can roll your own server, but often it's impossible.

The radar "reading" was done by first plotting analog radar signals to the antique rotary radar displays. Then there would be human operators with a light pen, marking each radar signature on each radar turn.

So the Univac would receive input coordinates for each target and track those in memory each turn.

CPU's microcode can be surprisingly simple: The CPU has bunch of internal signals, which activates certain parts of the CPU and the logic when to turn each signal on comes from reading bunch of input signals. The microcode can be just a memory where the input signals are the memory address and the output is the control signals.

Not hard, but time consuming. In the past two weeks I've had Claude Code write me around 35k lines of code across 350 commits. It's a project which is giving positive impact to the company, but we would never have started it without CC as the effort would have been too big compared to the impact.

I'm an avid user of the Claude Code planning feature and I like how Claude Code asks for questions. I also often iterate the plan before finally giving it a go.

How do you solve this in Kelos?

I tried to check the code base, but it didn't really provide any glues. I guess I could instruct the agent to build a plan and to post the plan in the issue and then iterate that with written comments in the issue. Is that how you run it?

There are companies which are only serving open weight models and not doing any training, so they must be profitable? Check for example this list https://openrouter.ai/meta-llama/llama-3.3-70b-instruct/prov...

But there are companies which are only serving open weight models via APIs (ie. they are not doing any training), so they must be profitable? here's one list of providers from OpenRouter serving LLama 3.3 70B: https://openrouter.ai/meta-llama/llama-3.3-70b-instruct/prov...

Not sure about that. My Android warns me about my wife's airtags so often, that if I would actually be tracked by a malicious airtag, I would just assume it's one of my wife's tags. This could be prevented if I could mark a tag to be trusted on my Android phone, but no such feature exists.

It's a Lenovo Yoga 9i with an Intel EVO i5 CPU. Not sure how much memory it has.

I was just helping my dad with a brand new Lenovo laptop with Windows 11. It felt unbelievable slow and sluggish. Just opening file manager to create a new folder lagged so much it felt like this would have been a 15 years old computer.

You can still do that as well

Hopefully you can write the teased next article about how Feedforward and Output layers work. The article was super helpful for me to get better understanding on how LLM GPTs work!

Thank you!

I remember having this argument with my professor at the school, who insisted that a function should have only one "return" clause at the very end. Even as I tried, I could not get him to explain why this would be valuable and how does this produce better code, so I'm interested on hearing your take on this?