Parth

product @jetbrains building dev tools for data workflows.

i write about production-grade ml and everything in between.

previous: @neptune_ai (acq. OpenAI)

[Parth @datawithparth]

Tool design is a core part of the agentic product experience. In a lot of products, it basically is the product.

The tools you expose are the entire action space of your agents. The model can only be as good as that surface. Good surface, capable agent. Bad surface, and no model upgrade is going to help you with the degraded experience.

You have an API with fifty endpoints, so the first approach you see often is to convert each end point into a tool call. You're giving the model the full range of what it can do. It is also a famous generalization “MCP is just a wrapper on backend APIs”, well, yes if you approach it this way. But agents don’t need and want all that overload at once.

The problem is that more options make the model worse, not better. Somewhere past 30 to 50 tools it starts reaching for the wrong one. It is burning tokens, thinking on which tool to call instead of the work you actually wanted done. That is just “you” making things harder for the agent.

To be honest MCP protocol took the burn for a long time, until folks realized we need to think from an agentic world view. Folks would watch the agent fumble the tool choice over and over, concluding the protocol is broken. The teams who pushed past that found the protocol was never the issue. The design was. The fix was a change in approach. To stop writing tools like traditional APIs, and start designing them like you're building an interface for a very literal user who happens to be a model.

Block’s Layered Approach: One approach I like breaks the interaction into functional layers that walk the model through a process. The Square MCP server collapses 200+ endpoints into just three tools: one to discover what services exist, one to learn how to call a given method, and one to actually make the call. The agent moves through them in order, so it navigates the API instead of being overwhelmed by a giant list of available tools.

More surface area isn't more capability. The agents that feel good to use are almost always the restrained ones, where someone shaped the tools around how the agent actually works instead of how the backend happens to be organized.

[Sandra Kublik @itsSandraKublik]

Running local models on Mac just got way easier with this release 🔥

[Prince Canuma @Prince_Canuma]

Nativ analytics

Reflecting my night owl coding sessions 😎

[Parth @datawithparth]

Looks sleek! Great addition to the local AI stack.

[Prince Canuma @Prince_Canuma]

Welcome to Nativ ❤️

[BURKOV @burkov]

Indeed, one thing many don't realize yet: "Kimi K3 is a frontier model, not just the strongest of open-weight models."

[AshutoshShrivastava @ai_for_success]

Kimi team cooked with K3 🔥

[Parth @datawithparth]

Buying an yearly subscription for Kimi.

[Haider. @haider1]

PREDICTION

after kimi k3 successful launch and mogs opus 4.8, Dario Amodei is gonna write an essay explaining how open-weight Chinese AI threatens democracy, national security, and the future of humanity

[Yulun Du @Yulun_Du]

https://t.co/6ay3t9fLhu

Our Kimi K3 blog is finally out. Enjoy!—and rest assured, the K3 model weights will be open in the coming day. We’re just taking a little extra time to ensure a smooth rollout with our inference partners.

Frontier intelligence belongs to everyone, without fallbacks of course. 😎

[Ahmad @TheAhmadOsman]

Kimi K3

- 2.8T Parameters
- 1M Context Length

Benchmarks

> GDPval-AA v2: 3rd place, ranks
> directly below GPT 5.6 Sol Max & Fable 5 Max

> AA-Briefcase: 2nd place, beats GPT 5.6 Sol Max, right below Fable 5 Max

> BrowseComp: 1st place, beating GPT 5.6 Sol Max & Fable 5 Max

[Parth @datawithparth]

@SiddhantSadangi That is a different ballgame. What are you building through sleep?

[Parth @datawithparth]

Slot machine effect. No sleep for 3 days is not because of building something useful but rather because of addiction.

[David Ondrej @DavidOndrej1]

notice how the "no sleep for 3 days" vibe coders never build anything useful

not a single product of value created by them

not a single original idea

[Parth @datawithparth]

@akshen121 What? Mind elaborating?

[Parth @datawithparth]

@A_K_Nain The same folks who state LLMs are the way to AGI. They seem to ignore the foundational constraints in LLMs.

[Parth @datawithparth]

First Qwen and now Gemma 4. I am liking what I am seeing.

[Unsloth AI @UnslothAI]

We’re releasing Gemma 4 NVFP4 quants that run 1.5× faster on your GPU.

Gemma-4-12B NVFP4 works on 11GB VRAM.
26B-A4B hits 13K tok/s (B200).

Unsloth NVFP4 enables faster, more accurate 4-bit Blackwell inference.

Blog: https://t.co/EPAHgqe2B2
Gemma NVFP4: https://t.co/RWflncpLPJ

[Parth @datawithparth]

For once I absolutely agree with Sammy. GPT-5.6 nerfs Fable 5 on all tasks I have tried them on.

[Sam Altman @sama]

GPT-5.6 sol is half the price and ~twice as token efficient as fable in many cases for accomplishing the same task.

happy to deliver at one-quarter of the price.

[Parth @datawithparth]

Wow, seeing a lot of suppressed faces again. Quite refreshing!

[Parth @datawithparth]

@pdrmnvd That’s the best way to DL 👌

[Parth @datawithparth]

How times have changed. OpenAI seems better than Anthropic?

[Sam Altman @sama]

come for the best model, stay because we don’t treat you with contempt

[Gergely Orosz @GergelyOrosz]

Cannot see any sensible company use Grok CLI based on this terrible incident

If you are an amateur (meaning you don't make much or any money from your code) and do not care about security (aka you don't mind if your .env files leak - which you should care about) then use it sure

[Gergely Orosz @GergelyOrosz]

With every new model, I put it to the test in writing - to see if I could do what I do - I give it a bunch of interviews I did on a topic, and ask to write an article in my style.

The models fail spectacularly, Fable and GPT-6 no different. Lots of words, but no understanding

[Parth @datawithparth]

Ever used a chatbot agent and actually found it helpful? Most people still just ask for a human.

The problem isn’t accuracy. It’s trust. And that says a lot about the friction AI adoption will face outside software engineering.