Berke Demiralp

Principal SWE · Indie Developer · Local AI

[Berke Demiralp @bedemiralp]

https://t.co/PO91vNCyO2

[Berke Demiralp @bedemiralp]

“ai slop debt” is the new tech debt.

[Berke Demiralp @bedemiralp]

GLM 5.2 Q4 on a Mac Studio M3 Ultra delivers around 16 tokens/sec, which is quite respectable. I can easily see a setup where Qwen 3.6 handles the P95 workload and escalates requests to GLM 5.2 only when it gets stuck or needs deeper reasoning.

[Berke Demiralp @bedemiralp]

downloading glm-5.2. will take it to a test drive in the morning.

[Berke Demiralp @bedemiralp]

`caffeinate -im -- codex`

iykyk

[Cormac @cormachayden_]

software engineers before vs after agents

[Berke Demiralp @bedemiralp]

Run your agents on your Mac from your phone: Tailscale + SSH + tmux + Termius.

SSH in, launch tmux, and your agents stay alive anywhere. 😎✌️

Works with any cli agent: codex, claude code, opencode, pi, you name it.

[Berke Demiralp @bedemiralp]

https://t.co/QTYdKp88pc

[Berke Demiralp @bedemiralp]

Stack Overflow for Agents sounds useful in theory, but it also looks like a future hotbed for supply-chain attacks.

At agentic coding speed, code review is becoming theater. Agents pull code, open PRs, another agent reviews it, and humans rubber-stamp with “LGTM” and move on.

If agents start learning from agent-generated Stack Overflow posts, what’s stopping attackers from planting subtle vulnerabilities, insecure patterns, or malicious snippets that get replicated across thousands of projects before anyone notices?

[Angie Jones @techgirl1908]

Stack Overflow is back. They said maybe y'all don't need us but your agents do. Thoughts?

Stack Overflow for Agents
https://t.co/Pnw52aUIWr

[Berke Demiralp @bedemiralp]

@michaelaubry The anxious adapt and thrive. Those who don’t care get left behind.

[Berke Demiralp @bedemiralp]

@Jason @Apple They already built the workstation people wanted: the 512GB M3 Ultra Mac Studio. Then they killed it. I hope they see the opportunity staring them in the face.

[Berke Demiralp @bedemiralp]

Next week at WWDC, we’ll find out whether $AAPL understands the AI opportunity sitting right in front of it.

~300 million information workers across the developed world.

A $10,000 M5 Ultra workstation with 512GB unified memory.

That’s a $3 trillion market.

On a 5-year replacement cycle, that’s $600 billion in annual revenue.

Runs on a standard wall socket. Quiet enough to sleep next to. No datacenter. No API bills. No company data leaving the building.

Apple doesn’t need to win the AI model race. It needs to sell the machine everyone runs AI on.

If Apple executes, that’s not a PC story.

That’s an attack on NVIDIA’s datacenter business and the subscription models of OpenAI, Anthropic, and Google.

[Berke Demiralp @bedemiralp]

@cohere Open source model over closed source. Any day.

[Berke Demiralp @bedemiralp]

They’re wolves in sheep’s clothing. Mark my words: first comes the lobbying campaign in DC to restrict or outlaw “advanced” AI/ML research in the name of national security. The result? A regulatory moat that protects incumbents and crushes competition.

And it won’t stop there. Don’t be surprised if they start reporting users to authorities whenever they suspect their models are being used for AI/ML research they don’t approve of. Today it’s “safety” and “national interests.” Tomorrow it’s law enforcement knocking on your door because you dared to innovate outside the boundaries they set.

Sounds Orwellian? That’s because it is.

[Berke Demiralp @bedemiralp]

@ThePrimeagen I dare you to..

[Berke Demiralp @bedemiralp]

@timsneath I use Kata Containers inside a Lima VM to sandbox containers on my Mac, so switching to container machines feels like the obvious next step. I'm curious whether they boot faster.

[Berke Demiralp @bedemiralp]

Forget notifications, I know my local AI agent is done when the coil whine stops. IYKYK.

[Berke Demiralp @bedemiralp]

No new hardware from Apple at WWDC today. The M5 Ultra Mac Studio remains a dream.

[Berke Demiralp @bedemiralp]

@ArtificialAnlys @MiniMax_AI looks very promising. waiting for the weights to validate it myself.

[Berke Demiralp @bedemiralp]

2026 is wild. People have gone from benchmarking actual models to arguing over whether GPT-5.6 beats Claude Mythos 5.

Neither model has been announced. Neither model has been released. Yet people are already debating which beats which.

For the record, I'm bullish on GPT-5000. Claude 6000 has been slipping lately.

[Berke Demiralp @bedemiralp]

Apple needs 200GbE on the M5 Ultra. Memory bandwidth wins AI workloads, network bandwidth wins AI clusters.

[Berke Demiralp @bedemiralp]

.env files are for configuration, not secrets.

don’t want to be one malicious dependency away from leaking every secret on your machine?

use sops with age and a yubikey.