Raymond Baghumian

Wordpress | Laravel Developer - MERN Stack

[Raymond Baghumian @rayiumir]

OpenClaw was one of the most overhyped AI products. Giving an AI access to your email, files, browser, and calendar is a serious privacy and security risk. One prompt injection or simple mistake can cause major trouble. Without a sandbox and strict limits, I don't use it.

[Raymond Baghumian @rayiumir]

The Claude Sonnet v5 model performs much worse in coding compared to DeepSeek v3.

[Raymond Baghumian @rayiumir]

Programming is by no means the first job in the world.

[Raymond Baghumian @rayiumir]

The Qwen3.8 community has optimized a list of GPUs for running AI models locally. For example, an RTX 4090 can reportedly go from around 36 to 75 tokens per second.

Given the current high cost of hardware, I absolutely do not recommend running local AI models.

[Sudo su @sudoingX]

the qwen 3.8 27b dense speed table, every number a paired a/b measured by the person who owns the card:

> rtx 5090 desktop: 66 → 144 tok/s, the current crown
> rtx 5090 desktop: 61 → 135, +120%
> rtx 4090: 36 → 75, +107%
> 3x 3090 tensor parallel: 49 → 96
> 2x 5060 ti, two flags stacked: 22 → 71, 3.14x total
> rtx 3090: 31 → 41 on day one, 41 → 63 on a fresh build
> rx 7900 xtx: 31 → 44
> radeon r9700: 27 → 47
> ryzen ai max apu: 11.5 → 23.7, doubled
> radeon 890m igpu: 2.7 → 5.7, doubled
> and dual 2016 tesla p40s sitting in the pr queue at +92%

three days ago this table didn't exist. 48 hours ago it had three rows. it now has 27 and it doesn't fit in a screenshot anymore.

six tuning rules found by the community, half of them by people i have never spoken to. nine more prs landing as i type this, the queue refills faster than i can merge it.

this is the command center for qwen 3.8 27b dense on real hardware. what are you waiting for?

[Raymond Baghumian @rayiumir]

Security and Test Writing; The Missing Link in AI-Assisted Programming

Artificial intelligence has made code generation several times faster, but the speed of code generation does not mean quality and security, and without test writing, 👇

[Raymond Baghumian @rayiumir]

Don’t even think about it, and don’t be tempted to run heavy AI models on your own hardware.

Getting just one token per second is simply not worth it. It puts a heavy load on your CPU, GPU, and other components, while also increasing power consumption and hardware wear. 👇

[Raymond Baghumian @rayiumir]

Connecting VSCode to Claude through the extension made it clear just how much RAM this combination can devour!

The VSCode editor, once known for its lightweight nature and reasonable resource usage, is now, with the addition of AI capabilities, 1/

[Raymond Baghumian @rayiumir]

If the claude model is left in its default state, it may perform much worse than solving a few lines of code, making the problem take longer to solve by requesting countless commands. #ai

[Raymond Baghumian @rayiumir]

@rezoundous Competition between companies allows costs to be controlled, subscription prices to increase less, and companies to offer more services and features to attract users.

[Raymond Baghumian @rayiumir]

@rezoundous $500? Even more...

Currently, the cost of developing and maintaining AI models is so high that even if there were no Chinese competitor, we would probably have to pay more than $1,000 for professional subscriptions in the next few years. 👇

[Raymond Baghumian @rayiumir]

Why we build software with frameworks like Electron or Tauri does not mean that it is not possible to write it with C++ or Rust; in fact, the native part of the software can be written entirely with Rust or C++. 1/

[Raymond Baghumian @rayiumir]

@gustavokov This is not good.

The problem with forking the kernel is that it will cause trouble in the Linux world later.

[Raymond Baghumian @rayiumir]

Yesterday, one of the managers said that Ford is hiring software engineers again because the company realized that overreliance on AI has created significant technical debt, making the cost of maintaining and fixing AI-generated code higher than hiring experienced developers. 1/

[Raymond Baghumian @rayiumir]

@PovilasKorop However, the problem is that the costs of advanced models are so high that it is not cost-effective for programmers who do not earn much.

[Raymond Baghumian @rayiumir]

@PovilasKorop The current situation of free AI models is not good, and companies have been trying hard to provide users with 50 to 100 free credits per month to encourage them to purchase the first plan and move towards advanced models.

[Raymond Baghumian @rayiumir]

@PovilasKorop Over-relying on AI won't develop real programming skills. It has unlimited tokens but the programmer does not have enough skills and does not learn anything. Unfortunately, illiterate programmers may become more common in the future.

[Raymond Baghumian @rayiumir]

@GithubProjects https://t.co/qidxrjtocw

[Raymond Baghumian @rayiumir]

Overusing AI alone won't make someone a programmer.

[Raymond Baghumian @rayiumir]

The situation of major AI companies in the United States and China shows that development and maintenance costs have increased significantly. This has led to serious discussions about token pricing and the profitability of AI models.

[Raymond Baghumian @rayiumir]

If Android becomes too restrictive and moves closer to a closed platform model, HarmonyOS could potentially evolve from a regional project into a global challenger.

[Raymond Baghumian @rayiumir]

Is Google Unintentionally Creating an Opportunity for HarmonyOS?

Android became popular over the years largely because of the freedom it offered users and developers. The ability to install apps from various sources, use alternative app stores, and enjoy a highly flexible 👇

[Raymond Baghumian @rayiumir]

That said, Google's greatest strength remains its vast ecosystem of apps and services. Yet the history of technology has repeatedly shown that dissatisfaction among users and developers can create opportunities for new competitors. 👇