CetinLM Trains 2B-Token AI Model on a Single Consumer GPU at Home

·2 views

Independent researcher Mert Cetin, working under Me Force Technology, has trained a language model called CetinLM past the 2.05 billion token milestone using only a single 16GB NVIDIA RTX consumer GPU in a home desktop setup. The project was built entirely from scratch, including custom data pipelines and a bespoke tokenizer, without relying on cloud infrastructure or enterprise hardware. Validation loss steadily improved from 2.7653 at 1.80 billion tokens to 2.7392 at 2.05 billion tokens, indicating stable and continuous learning. Even before any fine-tuning or chat alignment, the base model achieved a 75.1% top-token probability on factual recall tasks such as identifying Ankara as the capital of Türkiye. Cetin argues the project demonstrates that disciplined, efficiency-focused engineering can lower the barrier to AI research without requiring massive capital investment or large-scale compute infrastructure.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Experiment Shows External Memory Can Restore Learned Behavior After Full Brain Reset

A developer built a small recurrent neural network paired with an external memory system to test whether learned experience could survive a complete substrate reset. The experiment used a fixed cue-response task where the network had to map specific inputs to correct actions, with chance performance at 33%. Results showed that external memory alone achieved near-perfect accuracy, while plasticity alone underperformed at roughly 47%, and combining both yielded 88%. Crucially, the memory system was constrained to only bias internal network states rather than directly select actions, preserving the integrity of the test. The key finding was that externally stored experience could reinstate correct behavior in a reset network that had never independently learned the task.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Flock Safety Cameras Found to Contain Serious Security Flaws and Hardcoded Credentials

Security researcher Micah Lee has uncovered multiple vulnerabilities in Flock Safety's license plate recognition cameras, including hardcoded credentials embedded in the devices. These flaws could potentially allow unauthorized actors to gain access to the cameras or the data they collect. Flock Safety cameras are widely deployed by law enforcement agencies across the United States to monitor vehicle movements. The findings raise significant privacy and security concerns given the sensitive nature of the surveillance data these devices capture and store.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Installing Linux Packages Silently Runs Root Scripts You Never Audited

A software developer's routine installation of ffmpeg on a Linux machine prompted a deeper look at what package installation actually involves. Running apt install does far more than add a binary — it executes maintainer-written scripts as root, updates system databases, rewrites symlinks, drops config files into /etc, and triggers changes across dependencies. None of this process is sandboxed, meaning the user implicitly trusts the package maintainer and every upstream dependency with full system access. The author notes the greater risk lies not in obviously malicious packages but in routine, trusted ones that no longer attract scrutiny. The piece urges users and AI agents alike to treat any install command as a deliberate grant of root authority, not a harmless default action.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Critical Pre-Auth RCE Flaw Found in Fastjson 1.x; No Patch Available

A critical remote code execution vulnerability, tracked as CVE-2026-16723, has been disclosed in Fastjson versions 1.2.68 through 1.2.83, scoring as high as 9.8 on the CVSS scale. The flaw requires no authentication and exploits a trusted code path in the @JSONType annotation-handling branch, bypassing both AutoType and safeMode protections entirely. An attacker can host a malicious JAR file over plain HTTP and trigger execution simply by sending a crafted JSON payload to any endpoint calling JSON.parseObject(). Unlike previous Fastjson vulnerabilities, this exploit does not rely on known gadget classes or blacklisted type names, making existing defenses ineffective. Alibaba has declared Fastjson 1.x end-of-life and will not issue a patch, directing users to migrate to Fastjson2.

0 comments Read more at DEV Community