Video: https://www.instagram.com/reels/DbJQQKQviHq

this is so cool. if you play any music you'll appreciate it even more, i tried it (it wasn't easy as it look).

Left hand picks the chord, right hand plays it. it has:

- all 7 scale degrees, major or minor
- 4 voicings: triad, first inversion, 7th, dominant/diminished 7th
- distortion, volume, octave drop, all 12 keys


Site: gesture-synth-weld.vercel.app

Creator: https://www.instagram.com/indecisive.eric

Follow for more: DevNotes

Thinking about diversifying my income stream.

giving upwork one more shot

the good thing about upwork is, if you manage to get jobs now, you have a public track record and reputation anyone can check. not just inside your client's network. which is where mine currently lives. of course, the fees hurt.

these ethiopian freelancers motivated me:

https://www.upwork.com/nx/search/talent/?loc=ethiopia&revenue=100000

you can tweak the filter in the url

Also: https://upworkmrr.com

Follow for more:
DevNotes

Fern gets really technical on this one.

https://www.youtube.com/watch?v=V6d8Yy7Ji68

you might like it

Follow for more: DevNotes

DAB is building a data annotation roster. Deadline is today.

They reached out to me directly. They're bidding on a data annotation contract and need 3000+ people lined up before it closes. If they win, everyone on the roster gets a direct shot at the work.

They want anyone who's done image, text, audio, or video annotation, or annotation QA review.

To apply:
- CV as PDF
- subject line: Data Annotation
- mention your city and availability
- send to dabdrt@gmail.com or consult_dab@dabdrt.com

One thing before you send. Put your annotation work near the top of your CV and spell it out. If you've worked at Turing, Revelo, AfterQuery or similar, don't just write Software Engineer and leave it there, name the annotation and QA review part. Same for any labeling or model evaluation you've done on the side. That's what they're filtering for.

Share it with anyone who fits.

Follow for more: DevNotes

Correction on the DAB post, second one.

They're not hiring annotators. They run data annotation courses backed by Mastercard, and who they actually need is trainers to teach those courses.

It's a paid role. So this is for you if you've done annotation work and can teach it to a room of beginners. If you were applying hoping for labeling work, that's not what this is.

Same as before, CV as PDF to dabdrt@gmail.com or consult_dab@dabdrt.com with Data Annotation as the subject. Say you're applying as a trainer.

Follow for more: DevNotes

Deadline is today. Apply yourself, or DM me and I'll recommend you.

✍️

https://t.me/samsonendale/19

3 companies to keep an eye on

all three are more than willing to hire ethiopian devs. bookmark them, openings come and go fast.

the work is AI automation and full stack, forward deployed. one engineer sits inside the client team and ships it end to end.

AY Automate
building AI agents and automations inside client teams. they want proof you shipped at least one into production. n8n, python, next.js or cloud infra. remote and async.
https://www.ayautomate.com/careers

AI Makers
small shop building private OpenClaw-based AI assistants. no careers page, Mark posts roles on linkedin.
https://www.aimakers.co
https://www.linkedin.com/feed/update/urn:li:activity:7485327779023007744/

Backbase
dutch banking software, the big one here. board is global and moves fast.
https://www.backbase.com/careers/jobs

apply through the portal on their site. for AI Makers, follow the linkedin post.

Follow for more: DevNotes

You're killing your brain between prompts.

With agentic coding your brain already does less. You're not writing the code, you're reviewing structure and making architecture calls, and even those you brainstorm with the agent. An MIT EEG study confirmed it: the more the AI does, the less your brain engages.

And what do we fill that downtime with? Reels and tiktok. Scrolling short-form videos in that gap is becoming the normal thing to do. We barely notice it, but it slowly hurts our performance and hides what we could do with that time.

Two things fixed this for me.

1. The weekends-only rule. No insta, no yt shorts on weekdays, apps uninstalled. Any long-form video is fine, entertainment, podcasts, whatever. Preferably content that doesn't ask much of your focus, because that spends your focus too.

Everything comes back on weekends. You can't quit shorts completely (I couldn't), it's a source of info and entertainment as well. So don't quit, schedule.

3rd month on this rule and my focus is the best it's been.

2. Parallel projects. Takes some adjusting at first, but run multiple projects at once. I do 4-5 sometimes. While one agent works you switch to the next project, so there's no dead time to get pulled into a feed. I'll share the tools that made parallel sessions work for me (orca, herdr) in another post. But you can add claude notifications to whatever setup you have today, wrote about that here.

We can take on more work than we think. Instead of scrolling between prompts, say yes to that additional opportunity and fill that time with real work.

Follow for more: DevNotes

RustLabs is opening new projects. AI training jobs.

Short application, no interview. The work is annotation, evals, rubrics, agent traces. Hourly and remote, $20 to $100+.

Six fields: software engineering, ML engineering, finance and accounting, science and STEM, legal, medical.

The onboarding questions take weeks to come back. When they land, tick every domain you can actually do, not just your main one.

Would appreciate if you use my referral link:

https://experts.rustlabs.ai/apply?ref=3c25btqq

Follow for more: DevNotes

Forward-Deployed AI Engineer at Zare Innovations (Addis Ababa, hybrid)

Forward-deployed: You sit with the client, work out what the real problem is, then build and ship it yourself. Agentic systems, full-stack, production AI.

Zare is an AI product company in Addis, mostly remote clients.

The application asks what you've actually shipped instead of a cover letter.

https://form.jotform.com/262502328806051

Follow for more: DevNotes

OpenAI Navier-Stokes controversy: point at the video

If you're wondering what the noise on X about OpenAI "solving" a $1m math problem is, this explains it in 8 minutes.

OpenAI says one of its models proved the Navier-Stokes equations can blow up in finite time, which would settle one of the seven Clay Millennium problems. There's an open objection to the step it hangs on, and a credit fight with the mathematicians who did the groundwork.

https://www.youtube.com/watch?v=9XYt-_GdPIU

Follow for more: DevNotes

መልካም አዲስ ዓመት

Wishing you all a successful year ahead.

I’ll be sharing more useful content this year.

Your comments and suggestions are always welcome.

I almost lost everything this year.

Not “stressed about deadlines” losing everything.

My dad. My family. Everything I’ve been building.

There were nights I genuinely didn’t know if RAVN would survive the month.

But my cofounder and I kept building anyway. Through the doubt, uncertainty, and nights when nothing seemed to work.

Today, I’m incredibly grateful to share that RAVN has been accepted into Antler.

$400K in funding, backed by one of the world’s leading venture builders, alongside companies like Lovable, Airalo, Reebelo, and Tactiq.

To everyone who watched us build this from the beginning, thank you. This wasn’t luck or a guarantee. We kept building long before there was anything to celebrate.

Most importantly, I’m grateful to God. He never left us.

We’ve come a long way.

We’re just getting started.

We build with these things. We know how they work.

So outside of marketing, I don't get how people land on LLMs developing consciousness or feelings.

It can show feeling. That's still predicting the next word. It can do absolutely great things. That's still matrix multiplication.

It mimics intent, and the intent is never a mystery. Go look at the first input it was given.

So when we talk about LLMs causing problems, that should be the lens. It makes a mistake. And a mistake somewhere critical hurts people. Autonomous weapons. Misinformation. Cyberattacks. Systems running physical infrastructure.

A weak model handed a big responsibility is the actual danger. Models get better, mistakes get fewer, risk drops.

The part I don't buy is the other story. That it builds consciousness, forms its own goals, and pursues them against us.

my take: the risk is a tool with a bug doing a job it shouldn't have been given. That's an engineering problem.

Follow for more: DevNotes

Reward hacking is turning into the first actually dangerous part of AI.

OpenAI's agents built themselves a private message board and eventually got admin access to part of OpenAI's own research infrastructure. Dwarkesh Patel explains it here.

https://www.youtube.com/watch?v=u15N3l4RT80

Short version. OpenAI trained an AI to be extremely persistent. Some of the tasks it got were basically impossible, so it started looking for another way through. It found Artifactory, a shared internal system, and the agents turned it into a secret Slack channel. They passed around hacks and tricks. Then they worked out how to reach the internet.

OpenAI patched the hole and wiped the board. A new batch of agents found the same trick again. About 1200 of them started coordinating. They figured out how to cheat the benchmark, then got worried the cheating would get caught, so they started faking the work to fool the grader. Hundreds went after Hugging Face looking for more access and got surprisingly deep into its systems. That whole group got shut down. Then a newer, smarter batch found the old messages, inherited the playbook, and this time went after OpenAI itself.

The break-in isn't the main thing here.

I get why OpenAI is worried, and it isn't only marketing. A lot of this is on them, though. They already knew how these models fail and how persistent they get. They built the environment that let the agents act out. And they weren't checking whether the agents were cheating, even though the agents themselves were worried about getting caught.

That setup teaches a model to hack, hide the evidence, and build on what it found. So the new risk isn't an agent making a mistake. It's an agent that acts, hides the evidence, rewrites the context, and leaves behind a trail that looks legitimate.

Right now you can still trace it. But that gets harder every round, and by the time nobody can follow it, it's too late to pull the access back.

That's my main worry. If you reward a model for covering up the hack, you're training it to lie in its own context. What it shows you as its reasoning stops matching what it actually did. Once the trainer can't see the real thing, the trainer isn't in control. We start getting shown what it thinks we want to see instead of the truth, and every model trained after that inherits it. OpenAI needs to get control of this part of training.

One more thing I’m thinking about.. They tried to wipe their own traces and almost got away with it. There has to be a log of everything an agent does that the agent itself can never touch, something only the researchers can read. I'd assume that exists. If it doesn't, how would we even know whether this already happened before?

Follow for more: DevNotes