I gave OpenCode’s new Space Bunny model a task:
“Build me a playable 3D drone simulator.” 🚁
It one-shotted the core experience:
→ 3D environment
→ drone controls
→ flight physics
→ checkpoints
→ camera modes
→ weather
→ telemetry
→ responsive UI
Then I pushed it further and asked it to make the environment more cinematic
What surprised me most?
The speed
Usually, something like this means a lot of back-and-forth, debugging, and manual fixes
Here, OpenCode handled most of the build itself
And the result is actually playable in the browser. 💀
Space Bunny is a Flash-level model and the responsiveness really stood out while building this.
Available free on OpenCode for one week
Definitely testing it on some harder tasks next... 👀
2
3
9
902
You won’t land a job by just hitting “Apply” and tailoring your resume 100 times
That’s not a job search. That’s a waiting game
Try this instead:
→ Reach out to people on LinkedIn
→ DM hiring people on X
→ Ask for referrals
→ Reach out even for unpaid internships
→ Show what you can actually build
Be shameless about reaching out
One good conversation can do more than 100 applications
2
2
11
1,023
MLOps is easy to understand until you try to take a model to production
That’s why MLOps Zoomcamp by DataTalksClub is such a useful resource...👀
It takes you through the actual production lifecycle:
→ Experiment tracking + MLflow
→ Model registry
→ ML pipelines & orchestration
→ Batch, online & streaming deployment
→ Monitoring with Prometheus, Grafana & Evidently
→ Testing + CI/CD
→ Terraform
→ End-to-end MLOps project
And it’s free + hands-on
If you’re trying to go beyond “I trained a model” and actually learn how ML systems are deployed, monitored and maintained, this is one repo worth bookmarking
2
13
919
What if AI agents could learn to work without their harness...? 👀
Researchers from Google, Peking University, and HKUST just proposed Harness-Zero, a new approach to reducing an agent’s reliance on specialized harnesses
The idea is pretty interesting:
→ Train an agent with a specialized harness
→ Use a “harnessing agent” to turn the harness’s interventions into actions the model can learn
→ Fine-tune the model on those trajectories
→ Deploy it without the specialized harness
In their experiments, reported macro-average task success increased from 23.3% to 44.3% for Harness-Zero, compared with 41.7% for the base model with the specialized harness
They also report recovering 82.3% of harness-induced behaviors across 28 behavior patterns spanning knowledge work, tool use, and science
Basically:
Use the harness as training wheels → teach the model the behavior → take the training wheels off... 👀
8
2
16
1,085
Minko Gechev was preparing for a Google interview
Instead of keeping his prep notes to himself, he turned them into a GitHub repo
He collected the problems he actually practiced and organized them around the concepts he was preparing for
No “500 problems in 30 days”
No random problem dump
Just practical interview prep
He also openly points out that solving these problems isn't enough you need strong fundamentals and the ability to explain your thinking during the interview
That repo now has 3K+ stars. 👀
github.com/mgechev/google-in…
5
65
538
36,735
