Daily AI News

Baidu's Ernie 5.1 cuts 94% of pre‑training costs while competing with top models

Baidu unveiled Ernie 5.1, a compact LLM that uses a third of the parameters of its predecessor and costs only six percent of traditional pre‑training budgets, thanks to a "Once‑For‑All" sub‑model extraction technique. Early benchmark results show it rivaling the leading Western foundations models.

Read source →

OpenAI launches Daybreak, a security‑focused AI initiative

OpenAI announced Daybreak, a suite of AI agents that proactively scan codebases, model potential attack vectors and generate patches before vulnerabilities are exploited. The offering builds on the Codex Security AI launched in March and is positioned as a defensive layer for enterprise software supply chains.

Read source →

Meta and Stanford propose Fast Byte Latent Transformer to halve inference bandwidth

Researchers from Meta FAIR and Stanford introduced three inference methods for a Byte Latent Transformer that eliminate sub‑word tokenization and cut memory‑bandwidth consumption by more than 50%, potentially lowering the cost of serving large language models on edge hardware.

Read source →

Google stops zero‑day exploit developed with AI

Google’s Threat Intelligence Group detected and neutralized a sophisticated zero‑day vulnerability that threat actors had engineered using generative AI, marking the first public acknowledgment of AI‑assisted exploit development being foiled in real time.

Read source →

Nvidia signs $2.1 billion partnership with data‑center provider IREN

Nvidia entered a multi‑year $2.1 billion agreement with IREN to supply GPU accelerators and AI software stacks for European data‑center expansion, reinforcing Nvidia’s role as the backbone of next‑gen AI compute infrastructure.

Read source →

Nvidia has invested over $40 billion in AI partners in 2026

Continuing its aggressive ecosystem strategy, Nvidia disclosed that it has poured more than $40 billion into AI startups, cloud providers and research labs throughout 2026, cementing its position as the largest corporate backer of the AI industry.

Read source →

EU seeks direct regulator access to OpenAI’s GPT‑5.5 Cyber model, stalls on Anthropic

The European Commission is pressing OpenAI to grant regulators live access to its new GPT‑5.5 Cyber model for security review, while Anthropic has yet to comply, highlighting the regulatory tug‑of‑war over high‑risk AI transparency in Europe.

Read source →

Laserfiche rolls out AI agents for natural‑language workflow automation

Document‑management vendor Laserfiche introduced AI‑driven agents that execute workflow tasks based on plain‑language prompts while enforcing the platform’s built‑in security and compliance policies.

Read source →

Tilde Research unveils Aurora optimizer to fix hidden neuron‑death in Muon

Tilde Research released Aurora, a leverage‑aware optimizer that addresses a previously undocumented neuron‑death bug in the popular Muon optimizer, improving training stability and final model accuracy across a range of MLP architectures.

Read source →

OpenAI’s internal share sale creates roughly 75 multimillionaires

In a $6.6 billion internal share offering, about 75 OpenAI employees exercised the $30 million cap, becoming multimillionaires and highlighting the company’s growing valuation and employee wealth creation ahead of its upcoming public market plans.

Read source →