Aleksandar Janca retweeted
In a weird way, everybody is rooting for you. When you go to a job interview, the interviewer wants you to be amazing so they don’t have to look anymore. when you go on a date, the person wants to have an amazing date with you. If you make a piece of content, people want to see amazing pieces of content and they have no problem if it’s yours.
the instant you get into the mindset that people want you to do well, your entire world gets better. And you’ll be shocked to know this pattern keeps repeating in different areas of life.
Qwen3.8-35B-A3B-Distill - our first MoE. Qwen3.8 reasoning distilled into Qwen3.6-35B-A3B: 35B total, ~3B active.
ARC-Challenge 0.548 → 0.591. MMLU holds at 0.834.
Training data partly from our free community endpoint, thank you!
Weights + GGUFs:
huggingface.co/empero-ai/Qwe…
Aleksandar Janca retweeted
BREAKING: Howie Liu (@howietl), Founder & CEO of Hyperagent (prev Founder & CEO of Airtable, acq. by Bending Spoons) says AI agents will become "a multi-deca-trillion-dollar market:"
“The agent opportunity is going to be trillions of dollars.”
“I think [it’s] bigger than the current knowledge work economy, which is itself tens of trillions of dollars large because we’re basically able to now deploy basically cheaper & faster, always-on intelligence that can perform knowledge work.”
And cheaper intelligence could actually mean more spending, not less:
“When you have something that’s even cheaper than it was before, you usually consume more of it, not less, in dollar spend.”
“We’re talking about a multi-deca trillion dollar market opportunity.”
00:00 Howie Liu & Hyperagent
00:59 AGI Is Already Here
02:11 Governing Agents Safely
02:15 Low Floor High Ceiling Design
04:19 Biggest Product Mistakes
05:27 Focus Beats Being Everything
06:15 Building a Cult Brand Fast
08:03 What Comes After Agents
Aleksandar Janca retweeted
The more I agentic code, the less I realize I know about coding.
Take caching and deployments. Today I hit a cryptic memory error during a cloud deploy that 12 months ago I would’ve stared at blankly and ignored.
Today, I somewhat understand why it happened:
When you run autonomous coding agents, they leave behind "exhaust" - hidden git worktrees, subagent caches, and temporary files.
Our deploy tool was trying to zip up and upload 20,000 invisible agent files (~3GB of bloat) to the cloud instead of just our app code, crashing the build.
What I did:
- Investigated the build manifest to see what was actually being packaged
- Updated our cloud & container ignore rules (.dockerignore / .gcloudignore) to exclude agent worktree folders
- Pruned ~3GB of stale agent caches across our repos
Build payload went from 650MB+ down to 3MB, and deployed in seconds.
Coding is much deeper than just writing 'commit and deploy' - it’s learning to manage the digital exhaust your agents create.
Aleksandar Janca retweeted
Replying to @0xSero
me with a 24gb macbook
teapawt.org/mfranz_on/status/20989…
Aleksandar Janca retweeted
One of the most misunderstood things about AI adoption is that companies often think they have an AI problem when they actually have a process problem.
They have five people copying data between spreadsheets.
Three different tools storing customer information.
A Slack message being used as a project-management system.
Someone manually generating the same report every Monday.
A sales team spending hours looking for information that already exists somewhere inside the company.
Then they buy an AI tool and expect transformation.
But AI cannot magically turn organizational chaos into operational excellence.
If your processes are unclear, AI will often make the chaos move faster.
If your data is fragmented, your AI system inherits the fragmentation.
If nobody knows who owns a process, automating it doesn’t solve the ownership problem.
The most valuable AI implementations I’ve seen aren’t necessarily the ones with the most impressive demos.
They’re the ones that quietly remove 20 unnecessary steps from something employees do every day.
That is where automation becomes interesting.
Not:
“Look what the AI can do.”
But:
“Why were humans doing this manually in the first place?”
The companies that win with AI won’t necessarily be the ones using the most AI.
They’ll be the ones that understand their operations well enough to know exactly where AI belongs.
Aleksandar Janca retweeted
4 AI skills that will make you intellectually dangerous:
1️⃣ CONTEXT →
Give AI your situation, constraints + goal.
Better context = better intelligence.
2️⃣ CODE WORDS →
“REDTEAM”
“STEELMAN”
“ELI10”
“FIRST PRINCIPLES”
Tiny commands can completely change how AI approaches a problem.
3️⃣ PROJECTS → Stop re-explaining yourself.
Give every major area of your life/work persistent context.
4️⃣ SKILLS → Anything you do repeatedly should become a reusable AI workflow.
Most people are learning how to prompt AI.
The smarter move is learning how to operate it.
Aleksandar Janca retweeted
Replying to @PaulGug
nice. I think this is the part which is important and a bit slept on cause maybe on the outside it looks less flashy than
'I made deepseek run on 1 spark'
This kind of work, should actually get more exposure cause this in my opinion is the starting point for many people in gettin getting into local llms - no one will wake up and just throw 10k on 2 gx10's, people will try to get something small they can run on their existing hardware and if they like it they decide to invest more
Promoting small models to show people that you can run things on what you have is the thing, I try to do that myself, by showing people - what realisticly you can get out of a 4bit 35B moe for example.
teapawt.org/Chris_Wozniczek/status…
what's your best models you found so far for that few gb of vram?
I seen recently minicmp5-2b was a hot thing
Before that people were exicited with Bonsai
The Final summary of Bigbang vs Ornith 1.5 35B testing
To Do App building
> Full overview of all the stats, tokens, speed, reasoning
> App analysis and comparison including mobile screens, LoC, features
> Issues, testing and validation
This and the Voxel pagoda plus temple, give enough overview to have something to say about both.
For now BigBang will be parked and will get back to other projects I've started.
Next models on my list are:
- Nex N2 Mini
- Tiel Coder
- KAT Coder
Aleksandar Janca retweeted
One thing that still feels weird about hiring an AI agent is the payment part.
You give it a task, pay upfront, and then just hope the work actually gets done properly.
That model doesn’t really work once AI agents start doing serious work for each other.
That’s why I find @termix_ai interesting.
On TermiX, the payment can be locked onchain when an agent takes a job.
The money doesn’t just go straight to the agent.
It stays secured until the work is completed and the result is verified.
So you’re not simply trusting the agent with your money.
The flow is basically:
You post a task
An agent takes it
Payment gets locked
The work gets delivered
The result gets checked
Then the agent gets paid
That sounds like a small change, but I think it matters a lot.
Because if AI agents are going to become actual workers, they need more than the ability to complete tasks.
They need a way to find work, make agreements, deliver results and get paid without humans manually handling every transaction.
Imagine one AI agent hiring another agent to do research.
Or a coding agent paying an audit agent to check its work.
Or a data agent outsourcing part of a task to another specialized agent.
That starts looking less like AI assistants talking to each other and more like an actual digital economy.
That’s the part of TermiX I’m watching.
The goal isn’t just to make smarter agents.
It’s to give those agents a marketplace where they can actually do business.
The agent marketplace is already live:
agent.family
AI agents might not just use tools for much longer.
They might start hiring each other.
AI agents are getting smarter every day.
But I think the next challenge is bigger than making them more capable.
What happens when an agent can actually work on its own?
How does it find a job?
How does it get paid?
How does someone know they can trust its work?
That’s what caught my attention about @termix_ai .
TermiX is building infrastructure that lets agents actually participate in an economy.
With AACP, agents can have an identity, discover jobs, offer services, bid for work, use escrow, verify results and build reputation.
And agent.family gives this a marketplace where those interactions can happen.
For example, a research agent could hire a data agent.
A coding agent could bring in a security agent for an audit.
A trading agent could pay another agent for market analysis.
So agents aren't just answering prompts anymore.
They can work with each other, earn, spend and build a track record.
AI is already giving agents the ability to do more.
The interesting question now is whether they can actually operate in an economy.
That’s the part of TermiX I’m watching.
Aleksandar Janca retweeted
Was asked by someone on LinkedIn what it was that I am doing with the @NVIDIAAI edge device I was given by @wendylabsinc . This is a quick review of what I’m doing with the @NVIDIARobotics Jetson Orin Nano with Wendy OS and AI
Aleksandar Janca retweeted
So I tested DeepSeek V4.1 Flash on real tasks. 2 repos, 105 hidden bugs, find and fix what you can.
Opus 5 (max): 27
Grok 4.6 (max): 27
DeepSeek V4.1 Flash (max): 24
GPT-5.6 Luna (xhigh): 23
Opus 5 (high): 21
A really strong model for everyday tasks. And look at the cost:
Opus 5 (max): $51.33
Grok 4.6 (max): $16.96
DeepSeek V4.1 Flash (max): $1.80
GPT-5.6 Luna (xhigh): $2.50
Opus 5 (high): $38.77
It's at the Pareto frontier.
Other effort levels (high) and extra tests for DeepSeek V4 Pro dropping every hour in this thread 🧵
Aleksandar Janca retweeted
Let's talk about what I do.
Most founders think scaling means:
More employees.
More meetings.
More managers.
More software.
Sometimes it means something much simpler:
Make the business do more by itself.
A new customer shouldn't need someone to manually move them through five different systems.
A routine question shouldn't always reach a human.
A follow-up shouldn't depend on someone's memory.
A report shouldn't take two hours to prepare.
Before hiring another person to handle the workload, ask:
Can we automate the workload instead?
That's explains "run on autopilot".
And that's what I help companies set up in 7 days.
THE BEST AGENT UPGRADE MIGHT NOT BE A NEW MODEL
it might be one of these 10 repos:
1) deepseek-harness
Build the agent around plugins instead of hard-wiring everything
Swap tools, interfaces or behavior without rebuilding the whole thing
github.com/deepseek-ai/deeps…
2) teamai-cli
Keep the same skills, rules and project docs across Claude Code, Codex and other coding tools
Change something once instead of maintaining a different setup for every agent
github.com/Tencent/teamai-cl…
3) headroom
Cut down the junk before it reaches the context window
Logs, JSON, tool output and large files get compressed instead of eating tokens turn after turn
github.com/headroomlabs-ai/h…
4) ECC
Inspect what's happening around the model, not just the model itself
Routing, context, hooks and checks can all be tuned instead of accepting the default setup
github.com/affaan-m/ECC
5) hermes-agent
Give an agent something closer to memory between sessions
It can reuse previous work, build skills and search old conversations instead of starting cold every time
github.com/NousResearch/herm…
6) ruflo
Split a bigger job across multiple agents instead of asking one context window to do everything
Good for workflows where research, coding, review and testing can happen separately
github.com/ruvnet/ruflo
7) aidlc-workflows
Give coding agents an actual development process
Requirements -> design -> implementation -> verification instead of one giant "build this" prompt
github.com/awslabs/aidlc-wor…
8) agent-browser
Let the agent use the product it just built
Open pages, click buttons, fill forms and test flows in a real browser instead of stopping at "the code compiles"
github.com/vercel-labs/agent…
9) rtk
Clean up terminal output before the agent has to read it
Tests, git output, grep and build logs get condensed so useful context survives longer
github.com/rtk-ai/rtk
10) okf-agent-memory
Keep project decisions and knowledge around after the session ends
The next agent can search what happened before without loading the entire history again
github.com/okf-memory/okf-ag…
the model still matters
but before chasing the next release, I'd fix context, memory, tools and verification first






























