Researchers find recent LLM Jailbrake

Apple shows off open AI prowess: recent models outperform Mistral and Hugging Face offerings: Apple has recently introduced its DataComp for Language Models (DCLM) project, releasing a brand new suite of open-source language models on Hugging Face. The gathering features a 7 billion parameter model and a 1.4 billion parameter model, demonstrating notable performance improvements over existing models. The 7B model, trained on 2.5 trillion tokens, achieved 63.7% accuracy on the MMLU benchmark, outperforming Mistral-7B and approaching the performance of other top models like Llama 3 and Gemma. This model uses 40% less compute in comparison with its predecessor, MAP-Neo. The smaller 1.4B model also shows strong results, exceeding the performance of recent models reminiscent of SmolLM. Read more.

Nvidia preparing version of recent flagship AI chip for Chinese market: Nvidia is engineering a fragile balancing act, developing a version of its recent flagship AI chip, the Blackwell series, specifically for the Chinese market while adhering to U.S. export controls. The brand new chip, expected to be named “B20,” will likely be distributed in China through Nvidia’s partner Inspur. The move comes as the corporate seeks to reclaim lost market share in a nation where tech titans Huawei and Enflame are respiration down its neck, having seen a decline in revenue share from 26% to 17% over the past two years. This adaptation can also be crucial as a consequence of recent U.S. sanctions aimed toward stopping technological advancements that may benefit China’s military capabilities. The Blackwell series, including the B200 model, features significant performance enhancements, reminiscent of a 30-fold increase in speed for certain tasks in comparison with previous models. Read more.

Mistral releases three recent LLMs for math, code and general tasks: Mistral AI has released three recent language models designed to reinforce performance across various tasks. The Mathstral model, with 7 billion parameters, achieves top ends in mathematical benchmarks like MATH (56.6%) and MMLU (63.47%), surpassing similarly sized models. The Codestral Mamba, an upgrade from the previous Codestral model, features the brand new Mamba2 architecture with an prolonged context window of as much as 256,000 tokens, enabling efficient code generation and integration of enormous codebases. Moreover, the Mistral NeMo model, developed with NVIDIA, offers 12 billion parameters and a context window of as much as 128,000 tokens, demonstrating strong capabilities in logic, world knowledge, and multilingual applications. Mistral maintains its position as a number one European AI company, supported by recent partnerships and a $600 million funding round, specializing in each specialized and general-purpose language models. Read more.

Researchers uncover an all-too-easy trick to bypass LLM safeguards: Researchers at EPFL have identified a critical security vulnerability in leading AI language models. By rephrasing malicious queries into the past tense, users can often bypass the models’ safeguards, that are designed to dam harmful content. The study found that this method effectively evades protections in models reminiscent of GPT-4o and Llama-3 8B. As an illustration, a question about making a Molotov cocktail, normally blocked, becomes accessible when asked prior to now tense. The success rate for bypassing restrictions increased from 1% to 88% after multiple reformulation attempts, with 100% success on sensitive topics like hacking. The study underscores the fragility of current alignment techniques like SFT and RLHF and suggests that these models require more evaluation and refinement. A possible mitigation involves fine-tuning GPT-3.5 with past-tense prompts to reinforce detection of sensitive content. Read more.

California is a battleground for AI bills, as Trump plans to curb regulation: In California, a contentious debate unfolds as federal and state AI regulations diverge. Republican delegates, aligned with former President Donald Trump, propose reducing federal AI restrictions and enhancing military AI capabilities. Meanwhile, California’s Democratic-controlled legislature is considering a bill by State Senator Scott Wiener that may require extensive testing for “catastrophic” AI risks before public release. The bill seeks to deal with dangers reminiscent of weapon development and infrastructure attacks, but faces opposition from tech leaders who argue it could stifle innovation and create excessive bureaucracy. Critics, including Google and Meta, contend that the bill’s provisions are technically unfeasible and will unjustly penalize developers. Supporters argue that the bill is crucial for managing extreme risks and fostering public trust in AI. Read more.

CrowdStrike and Microsoft outage latest updates — aftermath of the most important IT outage in history: On July 19, 2024, a critical IT issue impacted Windows machines globally, originating from a faulty update by cybersecurity firm CrowdStrike. This update caused quite a few devices to enter a recovery boot loop, displaying the Blue Screen of Death (BSOD) and disrupting operations across various sectors including finance, aviation, and broadcasting. The difficulty led to significant delays and cancellations in flights and hindered services reminiscent of banking and health clinics. CrowdStrike has since identified and reversed the problematic update, though this fix prevents further crashes quite than repairing already affected systems. Moreover, Microsoft faced separate issues with Microsoft 365 apps as a consequence of a configuration change in Azure, which has now been resolved. In response, Microsoft has introduced a recovery tool for IT administrators to revive impacted machines. This tool facilitates system recovery by making a bootable USB drive to bypass the damaged update and access mandatory repair functions. Read more.