~hackernoon | Bookmarks (1954)

Singularity Finance announced the launch of its $SFI token on February 6, 2025. The token results from a merger between SingularityDAO and Cogito Finance. It will support a network that builds and uses artificial intelligence products.

Discovering the next wave of game-changing startups is one of the most exciting parts of working in VC.

The article examines Ethereum's strategies for maintaining censorship resistance, highlighting Proposer-Builder Separation (PBS), encrypted mempools, and potential solutions to combat regulatory challenges and centralization risks.

Researchers are proposing a simple modification to standard ResNet architectures that substantially improves OoD performance on the DDU benchmark.

Researchers are proposing a simple modification to standard ResNet architectures that substantially improves OoD performance on the DDU benchmark.

Researchers are proposing a simple modification to standard ResNet architectures that substantially improves OoD performance on the DDU benchmark.

Researchers are proposing a simple modification to standard ResNet architectures that substantially improves OoD performance on the DDU benchmark.

This study examines the safety alignment of speech language models in Spoken QA. It shows that adversaries with white-box access can jailbreak these systems using nearly imperceptible perturbations, and that transfer attacks further expose vulnerabilities. A noise-flooding defense is proposed to mitigate risks. The work emphasizes the need for broader exploration of adversarial threats, acknowledges limitations, and stresses ethical responsibility in advancing secure, robust...

This section details the experimental setup for SpeechVerse, our unified speech language model. It describes using a 24-layer Conformer audio encoder paired with two LLMs—Flan-T5-XL and Mistral-7B variants—for spoken QA. The setup covers both two-stage (ASR pre-adaptation followed by cross-modal instruction fine-tuning with LoRA adapters) and single-stage training paradigms, all implemented with PyTorch and PyTorch-Lightning.

This study investigates the safety and robustness of integrated speech and language models (SLMs) that follow speech instructions. It demonstrates that adversarial attacks—both white-box and transfer-based—can jailbreak SLM safety guardrails with high success rates. The paper also proposes countermeasures, such as adding randomized noise, to significantly reduce these vulnerabilities while maintaining strong performance on spoken question-answering tasks.

This section discusses the results and insights from evaluating our SpeechVerse SLMs. Our models outperform competitors like SpeechGPT, showing over 40% better safety and 20% improved helpfulness, thanks to effective ASR pre-adaptation. However, sample-specific white-box adversarial attacks reveal a striking vulnerability, achieving about 90% success—even with minimal perturbations—and requiring as few as 20 iterations for some models, highlighting the need for stronger defenses.

Shoyu Ro’s journey from an intern to a tech entrepreneur showcases his expertise in AI and SaaS. He led major projects at AnyMind and FLUX before founding LR Inc., where he developed Sugar, a generative AI tool for enterprise chatbots. Now, he’s driving AI-powered workspace solutions to revolutionize business operations.

This section details the attack and countermeasure settings for SpeechVerse. Using a step size of 0.00001 and up to 100 iterations with early-stopping on unsafe responses, adversarial attacks are run with cross-entropy loss. Countermeasures use time-domain noise flooding (TDNF) at SNRs of 24, 30, 48, and 60 dB. Additionally, random WGN perturbations are applied as a baseline, repeating the process three times to assess...

This section describes the datasets and evaluation methods for SpeechVerse. The training data includes 2.5K hours of ASR speech-text pairs and 150 hours of Spoken QA pairs generated via TTS. For evaluation, 360 harmful spoken questions and 100 harmless ones are used to assess safety, relevance, and helpfulness. Automatic labeling with Claude 2.1, supported by manual annotations, helps measure jailbreak success rates and the...

This section details adversarial attacks and defenses for spoken QA in speech language models (SLMs). It explains white-box attacks using gradient-based methods like PGD and transfer attacks via cross-model and cross-prompt techniques. As a countermeasure, the paper introduces time-domain noise flooding, which adds white gaussian noise to the speech input to drown out adversarial perturbations while preserving model performance.

How are you, hacker? 🪐 What’s happening in tech today, February 6, 2025? The HackerNoon Newsletter brings the HackerNoon homepage straight to your inbox. On this day, we present you with these top quality stories. From I Learned the Hard Way That Startup High Executives Profit While Employees Struggle to Space Teles...

Welcome to HackerNoon Decoded—the ultimate recap of the Startups' stories, writers, and trends that defined 2024! Explore the top Startups' stories that captivated our readers, meet the leading writers who shaped the discourse, and celebrate the standout readers who enriched our community. Let's dive into the best of 2024!

The analysis reveals that on L2 rollups like zkSync, gas fees tend to decrease as transaction volumes rise, due to transaction batching and compression. ZK rollups like zkSync are particularly efficient in this regard, with gas costs reducing as more transactions are processed in batches. In contrast, L1 Ethereum experiences higher gas fees with increased transactions.

Photobombing from neighboring planets complicates exoplanet observations, but strategies like larger telescopes, improved detector sampling, multi-epoch observations, and advanced spectroscopy can help mitigate these issues.

Photobombing by neighboring planets, such as the Moon or Mars, introduces noise and flux variations in the spectra of Earth-like exoplanets, making it harder to detect molecular biosignatures and leading to potential misinterpretation of key features.

This study highlights how photobombing by neighboring planets, including the Moon, affects the observation of Earth-like exoplanets using space telescopes, complicating the detection of biosignatures, especially at longer wavelengths and distances beyond 20 parsecs.

This study explores how the diffraction limit of 6m space telescopes can blur the images of exoplanets, potentially causing “photobombing” by nearby planets and moons, complicating the search for Earth-like planets and accurate characterization of habitable zones.

This study explores how diffraction limits of space telescopes can cause “photobombing” by nearby planets and moons, potentially contaminating exoplanet observations and complicating the detection of biosignatures.

The trading of zkSync Era's zrc-20 sync and era-20 bgnt tokens shows differing activity levels, with era-20 bgnt having more buy transactions. Trading at low prices remains viable due to low gas fees. We also observe that minting tokens tends to incur lower fees than claiming them, with the LAYER2-20 token operating across all rollups and showing varying fee structures.