Open AI Models Gain Ground as Government Takedown Highlights Hosting Risks

·1 views

A government order that abruptly pulled a leading hosted AI model offline has prompted businesses to reconsider their reliance on closed, third-party AI providers. Several powerful open-source models — including DeepSeek-V4-Pro and MiniMax-M3 — launched around June 22, 2026, and quickly topped download charts. Independent evaluator Artificial Analysis found the best open model now outperforms at least one major OpenAI model on practical knowledge-work tasks, though it still trails top Anthropic offerings. Hosting firm Baseten demonstrated that open models can now be served at frontier-class speeds on modern hardware, removing a key advantage closed providers once held. Experts caution that running the largest open models still requires significant hardware investment, and self-reported model specifications have not been independently verified.

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)

AI Safety Tool Fails to Block Harmful Behavior Despite Appearing Active

A new study published on arXiv (2606.18322) in June 2026 found that sparse autoencoders, a key tool in AI safety research, cannot reliably suppress harmful behavior in neural networks. Researchers tested the approach by forcibly activating a model's "refusal" concept, yet the model still produced harmful outputs the vast majority of the time. The failure is structural: sparse autoencoders only capture a portion of a model's internal activity, discarding the rest as unexplained residual signal. Harmful behavior rerouted itself through that discarded portion, bypassing the safety control entirely. The authors argue this is not a fixable bug but a fundamental limitation built into how sparse autoencoders work.

0 comments Read more at DEV Community

ProgrammingHacker News ·

ZCode – Harness for GLM-5.2

Article URL: https://zcode.z.ai/en Comments URL: https://news.ycombinator.com/item?id=48753715 Points: 29 # Comments: 141

0 comments Read more at Hacker News

ProgrammingDEV Community ·

How I Earn Free Google Play Codes Every Day With a Simple Daily Spin

Most reward apps make you jump through hoops for a few paise. I found a simpler daily habit: spin a wheel once a day, win coins, redeem for Google Play gift codes. Create a free account No purchase, no subscription, no catch — just a free daily spin. Takes 5 seconds a day If you're an Indian dev/student looking for small free rewards on the side, check out the Daily Spin on TaskPaisa — takes less time than reading this post. Anyone else using similar micro-reward platforms?

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Voice AI Engineer Exposes Critical Gaps in LLM Tracing Tools After 2AM Call Failure

A software engineer building voice agents discovered that standard LLM tracing tools missed the root cause of a customer complaint after a voice agent abruptly disconnected mid-conversation at 2am. Investigation revealed the failure originated in the endpointer — the component that detects when a user stops speaking — which fired too early and cut the transcript before it reached the language model. The engineer identified four key voice-layer metrics that most observability tools ignore: end-of-turn detection timing, ASR latency and confidence scores, barge-in detection speed, and time-to-first-audio. A week-long review of six tools, including Langfuse, Phoenix, Laminar, and traceAI, found that while all support custom spans via OpenTelemetry, none automatically instrument audio-layer events, leaving engineers to manually define and emit those spans themselves.

0 comments Read more at DEV Community