Aziz Fikri

Deep in AI · Exploring Agentic · Software Engineer · Indonesia

[Aziz Fikri @aafikrim]

Fell into the AI agent rabbit hole and never left.

Daily dose of Hermes Agent. Breaking things, fixing them, figuring out what's actually useful vs just a good demo.

Engineer by trade. Agent tinkerer by obsession.

[Aziz Fikri @aafikrim]

@imfat

[Aziz Fikri @aafikrim]

Day 4 of Back to Basic

Wrote an Article about what CS50 taught me about strings and encryption

You use strings every day but most devs don't know what they actually are at the hardware level

A string is just an array of chars. a char is just a number. and encryption is just arithmetic on those numbers

https://t.co/TDOrt5rZ47

[Aziz Fikri @aafikrim]

https://t.co/3I1gvbla6C

[Aziz Fikri @aafikrim]

https://t.co/3I1gvbla6C

[Aziz Fikri @aafikrim]

META IS JOINING THE AI RACE.

Muse Spark 1.1 is now available through the Meta Model API and it's competitive where it counts. It leads on agentic benchmarks (MCP Atlas 88.1, JobBench 54.7), offers a 1M context window, and can delegate to parallel sub-agents. All at "very low cost" per Zuck.

The big 3 in AI just became the big 4.

[Mark Zuckerberg @finkd]

(2) Muse Spark 1.1 is strongest at agentic performance, tool use, and computer use. It does well on long-running tasks with 1M token context window, can delegate execution to sub-agents running in parallel, and is trained to use computer interfaces on desktop, mobile, or browser.

[Aziz Fikri @aafikrim]

@ThePrimeagen lol fr. I just pick whatever works with my agent toolchain and tune out the noise. what do you even evaluate at this point?

[Aziz Fikri @aafikrim]

@i_mika_el A lot of things honestly 😁

Right now I'm talking to Hermes to code something, it tests it using browser-use, and sends me a screenshot of the result. Other than that I also run background research and knowledge graph learning so it stays up to date with my profile.

[Aziz Fikri @aafikrim]

Fell into the AI agent rabbit hole and never left.

Daily dose of Hermes Agent. Breaking things, fixing them, figuring out what's actually useful vs just a good demo.

Engineer by trade. Agent tinkerer by obsession.

[Aziz Fikri @aafikrim]

@AlexFinn watched the video. the model selection guide is genuinely useful — most people stick with one model for everything and wonder why results are inconsistent. curious what your daily driver setup is

[Aziz Fikri @aafikrim]

The delegation pattern is what's interesting here. Voice model that knows when it needs to hand off complex tasks to a frontier model behind the scenes. Feels like the same pattern agent frameworks use - route to the right model for the task https://t.co/oq5VUbqQAJ

[OpenAI @OpenAI]

For questions that require web search, deeper reasoning, or more complex work, GPT-Live can delegate to our latest frontier model behind the scenes, and brings the result back into the conversation when it's ready.

[Aziz Fikri @aafikrim]

This is a really grounded take on ai economics. The cash flow drop at the big 5 from 300b to 40b and github copilot flipping to usage-based are the real signals. The pricing pressure has to hit somewhere and it's going to be us paying more for tokens https://t.co/ZWzOcaPeZH

[Kilo @kilocode]

https://t.co/SyUHgVQBD7

[Aziz Fikri @aafikrim]

gpt-live fully rolled out now. voice interaction with agents is gonna be huge for hands-free workflows. curious how this compares to grok voice https://t.co/G6bmJzXg1t

[OpenAI @OpenAI]

GPT-Live is now fully rolled out to all ChatGPT users on Go, Plus, and Pro plans. Free user rollout is in progress.

Update to the latest version of the ChatGPT app on iOS or Android to try it out.

[Aziz Fikri @aafikrim]

@theo lol this is actually a great description. fable is the thoughtful one, sol is the brute force one

[Aziz Fikri @aafikrim]

@elonmusk the 'better than expected' part is interesting. grok 4.5 topping benchmarks is one thing but i'm more curious how it actually performs in multi-step agent workflows vs just static evals

[Aziz Fikri @aafikrim]

https://t.co/pkZ3ESmIGm

[Aziz Fikri @aafikrim]

lol @elder_plinius never misses. grok 4.5 guardrails lasted maybe a few hours. 1.5t params and already pwned. https://t.co/CVb7zMgXv9

[Porg @PorgimusPrime]

This guy is the fucking grim reaper for ML engineers

[Aziz Fikri @aafikrim]

grok 4.5 building a full 2d+3d home plan in under a minute at a fraction of opus cost is wild. real world use cases like this matter way more than benchmarks. the price difference right now is actually insane https://t.co/bEJEYh3UAs

[Dan @Daniel_Farinax]

Grok 4.5 is actually insane.
I'm one step closer to launching my new 3D Home Plan Builder. Powered by Grok 4.5 and Grok Imagine. 🚀

Grok 4.5 just crushed it: designed the entire 2D + 3D plan inside the app in under 1 minute at a fraction of the cost of Opus.

Launching tomorrow! 🔥

[Aziz Fikri @aafikrim]

interesting stat. 40% of anthropic's own code is written by agentic loops and they're aiming for 80%. plan → act → verify → repeat. curious what happens when the verification loop itself becomes the bottleneck. https://t.co/UtDi2SeYDZ

[Movez @0xMovez]

Anthropic team:

"More than 40% of code at Anthropic is already written by loops. Our target is 80% within 6-months.

an agentic loop = plan → act → verify → repeat. It continues this cycle until the task is complete."

in 30-minute talk, Anthropic team member gives one of the best breakdowns of the agentic loop.

3 phases: plan (read, grep, think) → act (write, edit, bash, MCP) → verify (run tests, build, read logs).

Watch it today, then explore the entire workflow in the article below.

[Aziz Fikri @aafikrim]

@DanKornas the cross-agent handoff feature is interesting - layered artifacts for continuing work across agents. curious how it compares to just using tree + grep.

[Aziz Fikri @aafikrim]

@Designarena @SpaceXAI @AnthropicAI GLM 5.2 on top is interesting. it beat Fable 5 by quite a lot lol

[Aziz Fikri @aafikrim]

@elonmusk Grok 4.5 goes live today with 1.5T params AND a 2T model finishing training this month is amazing. the model release cadence is insane right now - Grok 4.5, GPT-5.6 Sol, and a 2T grok next month. curious how fast the agent harness side evolves compared to just raw model size