Hasty Briefs - Daily Tech News Brief

Bilingual

FurtherAI (YC W24) Is Hiring Software and AI Engineers
a year ago
- FurtherAI is developing AI Teammates to automate insurance workflows.
- AI Teammates can handle unstructured documents, data entry, and phone calls.
- The company's mission is to make AI as reliable and adaptive as human employees.
- Successfully raised pre-seed funding from Y Combinator, South Park Commons, and Converge VC.
- Founding team includes a second-time founder and an Apple language modeling scientist with 12 years of collaboration.
Show HN: Dia, an open-weights TTS model for generating realistic dialogue
a year ago
- Dia is a 1.6B parameter text-to-speech model by Nari Labs, generating realistic dialogue from transcripts.
- Features include emotion/tone control, nonverbal sounds (laughter, coughing), and audio conditioning.
- Pretrained model checkpoints and inference code are available on Hugging Face.
- Demo page compares Dia to ElevenLabs Studio and Sesame CSM-1B.
- Community support via Discord; waitlist for larger model access.
- Installation via GitHub: clone repo, set up environment, and run Gradio UI.
- Python code example for generating dialogue audio with Dia.
- Supports GPUs (PyTorch 2.0+, CUDA 12.6); CPU support coming soon.
- Real-time audio generation on enterprise GPUs; slower on older GPUs.
- Full version requires ~10GB VRAM; quantized version planned.
- Strict usage restrictions: no identity misuse, deceptive content, or illegal activities.
- Future plans: Docker support, inference optimization, quantization.
- Team of 1 full-time and 1 part-time engineers; contributions welcome.
- Acknowledgments: Google TPU Research Cloud, SoundStorm, Parakeet, Descript Audio Codec.
Spark AI (YC W24) is hiring a full-stack engineer in San Francisco
a year ago
- Spark builds AI research tools to help energy developers navigate regulations and build solar farms and battery plants.
- Used by industry leaders like Colliers Engineering & Design, Standard Solar, and Cypress Creek Renewables.
- Customers' energy pipelines produce 60GW, powering tens of millions of households annually.
- Funded by top investors including AI Grant (Nat Friedman, Daniel Gross), Brex founders, and others.
- Team includes ex-Tesla, Brex, Google engineers; in-person work in San Francisco, 5 days a week.
- Tech stack includes Typescript, NextJS, NodeJS, and Postgres.
- Looking for engineers with 3+ years experience who value impact over perfection and want to learn the business side.
- Founded by Tae and Julia, former Tesla, Brex, and Apple leaders.
From 'catch up' to 'catch us': How Google took the lead in enterprise AI
a year ago
- Google has made a significant turnaround in the enterprise AI race, showcasing benchmark-topping models and a cohesive strategy at Google Cloud Next 2025.
- Google's Gemini 2.5 Pro is now considered the best-performing LLM, excelling in reasoning, coding, and offering a 1 million token context window.
- Google's infrastructure, including the new Ironwood TPU, provides unparalleled efficiency and performance, supporting its AI models and enterprise solutions.
- Google's integrated full-stack approach, combining models, infrastructure, and enterprise tools like Vertex AI and BigQuery, offers a competitive edge.
- Google's focus on AI agents and enterprise value is resonating, with 500+ customers in production and notable partnerships with companies like Wendy's and Salesforce.
Cekura (Formerly Vocera) (YC F24) Is Hiring
a year ago
- Cekura is a Y Combinator–backed startup focused on AI voice agent reliability.
- The company automates testing and observability for AI voice agents using realistic conversational scenarios.
- Cekura's platform uses AI-generated datasets, dynamic persona simulations, and real-time monitoring to optimize performance.
- The startup aims to reduce time-to-market and minimize production errors in the voice agent market.
- Cekura is looking for a Founding Engineer to lead development and collaborate with founders and customers.
'Immediate red flags': questions raised over 'expert' much quoted in UK press
a year ago
- Barbara Santini, an Oxford-educated psychologist, has been widely quoted as an expert in various topics.
- Concerns have been raised about her qualifications and identity, leading to the removal of her comments from major news outlets.
- Santini's main online presence is as a sex and relationships adviser at an online sex toy outlet, Peaches and Screams.
- The British Psychological Society (BPS) confirmed she is not a member, and her social media presence is minimal.
- Questions about Santini were first raised by the Press Gazette, and attempts to verify her identity were unsuccessful.
- Some reporters received her comments through companies connecting journalists with experts, leading to investigations by these platforms.
- Santini's profiles on expert-journalist platforms like Qwoted were removed due to red flags about her credentials.
- Several news outlets, including Reach and the BBC, have removed articles featuring her comments.
- The case highlights the challenges journalists face in verifying sources in the AI age.
- Experts like Charlie Beckett and Roman Raczka emphasize the need for better verification processes and professional membership checks.
Columbia student suspended over interview cheating tool raises $5.3M
a year ago
- 21-year-old Chungin 'Roy' Lee raised $5.3M in seed funding for his startup Cluely, an AI tool designed to 'cheat' on exams, sales calls, and job interviews.
- Cluely originated from a viral X thread where Lee was suspended from Columbia University for developing an AI tool to cheat on software engineering job interviews.
- The startup's tool, originally called Interview Coder, operates via a hidden in-browser window undetectable by interviewers or test givers.
- Cluely compares its AI tool to past inventions like calculators and spellcheck, which were initially criticized as 'cheating.'
- A controversial launch video showed Lee using Cluely's AI to lie on a date, drawing comparisons to 'Black Mirror.'
- Cluely has surpassed $3M in Annual Recurring Revenue (ARR) and was co-founded by Neel Shanmugam, another 21-year-old former Columbia student.
- Both co-founders dropped out of Columbia following disciplinary actions related to their AI tool.
- Cluely started as a tool to cheat on LeetCode, a coding platform the founders consider outdated.
- Lee claims to have secured an Amazon internship using Cluely, though Amazon prohibits unauthorized tools in interviews.
- Cluely is among several controversial AI startups launched recently, including one aiming to replace all human workers.
AI hallucinations lead to a new cyber threat: Slopsquatting
a year ago
- Researchers warn of 'Slopsquatting', a new supply chain attack exploiting AI-generated fake package recommendations.
- AI models like GPT-4, CodeLlama, and DeepSeek hallucinate non-existent packages, with 19.7% of recommendations being fake.
- Open-source AI models hallucinate more frequently (21.7%) compared to commercial ones (5.2%).
- Threat actors can register these hallucinated package names to distribute malicious code, posing widespread risks.
- Hallucinated packages are persistent (43% reappeared in tests) and semantically convincing (38% similar to real packages).
- Developers advised to use dependency scanners and avoid rushing security testing to mitigate risks.
Is SaaS a good business model for drug‑discovery companies?
a year ago
- The 2024 Nobel Prize in Chemistry recognized advances in computational protein design, sparking interest in foundation models for biology.
- Proteins play a crucial role in drug effectiveness, and AlphaFold2 demonstrated deep learning's potential in predicting protein structures.
- Companies like EvolutionaryScale and Profluent are developing powerful generative models for protein science.
- Licensing AI models as SaaS to pharma is considered a weak business model due to misalignment with pharma's core needs.
- Pharma's main challenge is the 'Valley of Death' between molecule discovery and clinical success, where current models fall short.
- Drug developers focus on ADME, PK/PD, and CMC, areas where foundation models currently lack reliability.
- The failure of Merck’s verubecestat highlights the gap between in-silico predictions and clinical outcomes.
- AI-first startups like Recursion and Insitro are pivoting to owning their molecules to control more of the value chain.
- Open-source models, like the replication of AlphaFold3, threaten the pricing power of SaaS licenses.
- Pharma prioritizes clinical success over molecule discovery costs, making SaaS models less attractive without downstream risk reduction.
I Open-Sourced My AI Toy Company That Runs on ESP32 and OpenAI Realtime API
a year ago
- Realtime AI Speech powered by OpenAI Realtime API, ESP32, Secure WebSockets, and Deno Edge Functions for >10-minute uninterrupted global conversations.
- Install Supabase CLI and set up your Local Supabase Backend.
- Set up your NextJS Frontend with environment variables and run the development server.
- Add your ESP32-S3 Device MAC Address to the Settings page in the NextJS Frontend.
- Add your OpenAI API Key in the server-deno/.env and frontend-nextjs/.env.local file.
- Start the Deno server with environment variables.
- Set up your ESP32 Arduino Client and configure the device wifi.
- ElatoAI consists of three main components: Frontend Client, Edge Server Functions, and ESP32 IoT Client.
- Features include Realtime Speech-to-Speech, Custom AI Agents, Customizable Voices, Secure WebSockets, Server VAD Turn Detection, Opus Audio Compression, Global Edge Performance, ESP32 Arduino Framework, Conversation History, Device Management, User Authentication, Conversations with WebRTC and Websockets, Volume Control, and Realtime Transcripts.
- Technologies used: Frontend (Next.js, Vercel), Backend (Supabase DB), Edge Functions (Deno Edge Functions on Deno/Supabase), IoT Client (PlatformIO, Arduino Framework, ESP32-S3), Audio Codec (Opus), Communication (Secure WebSockets), Libraries (ArduinoJson, WebSockets, AsyncWebServer, ESP32_Button, Arduino Audio Tools, ArduinoLibOpus).
- Performance metrics: Latency (<1s round-trip globally), Audio Quality (Opus codec at 24kbps), Uninterrupted Conversations (Up to 10 minutes), Global Availability (Optimized with edge computing with Deno).
- Security features: Secure WebSockets (WSS) for encrypted data transfers, Optional API Key encryption with 256-bit AES, Supabase DB for secure authentication, Supabase RLS for all tables.
- Limitations: 3-4s Cold start time while connecting to edge server, Limited to upto 10 minutes of uninterrupted conversations, Edge server stops when wall clock time is exceeded, No speech interruption detection on ESP32.
- Future improvements: Looking for Speech Interruption detection on ESP32, Adding Arduino IDE support, Adding tool calling support on Deno Edge.
- Contribution guidelines: Fork the repository, Create your feature branch, Commit your changes, Push to the branch, Open a PR.
- Project is licensed under the MIT License.
π0.5: A VLA with open-world generalization
a year ago
- Robots have advanced significantly, performing complex tasks like folding laundry and cleaning tables.
- The biggest challenge in robotics is generalization—adapting to new settings and objects.
- Generalization requires robust physical skills and common-sense understanding of the environment.
- Most commercial robots operate in controlled environments like factories due to limited generalization.
- Robotic foundation models like π0.5 aim to generalize to messy, real-world environments.
- π0.5 can perform tasks in new homes, showing flexibility and resourcefulness.
- Co-training on heterogeneous data enables π0.5 to understand semantic context and transfer skills.
- π0.5 combines high-level semantic decisions with low-level motor control for complex tasks.
- Experiments show π0.5 can clean kitchens and bedrooms in unseen environments.
- Future improvements include autonomous learning and better knowledge transfer.
More than 20 robots race alongside humans at half marathon in Beijing
a year ago
- More than 20 robots participated in a half marathon in Beijing alongside human runners.
- Robots varied in size, with some as short as 120 cm and others as tall as 1.8 meters.
- Tiangong Ultra, the winning robot, completed the race in 2 hours and 40 minutes, aided by long legs and a specialized algorithm.
- Some robots struggled during the race, with incidents like falling at the start or crashing into railings.
- This event marked the first time humanoid robots raced alongside humans in a 21-kilometer marathon.
- China views robotics as a frontier industry that could drive future economic growth.
- Experts question whether marathon participation is a reliable indicator of robots' industrial potential.
Onyx (YC W24) Is Hiring for ML Engineer
a year ago
- Onyx is an open-source AI assistant and enterprise search platform with a large user base.
- The project has over 10K stars and 3K+ community members on Slack and Discord.
- Onyx acts as a knowledge layer on top of LLMs, improving agent and knowledge retrieval capabilities.
- It syncs data from various sources (Google Drive, Slack, GitHub, etc.) to provide centralized knowledge access.
- Onyx aims to bring knowledge-enhanced GenAI to all teams within the next 5 years.
- The company raised a $10M seed round backed by Khosla Ventures and First Round Capital.
- Customers include Netflix, Ramp, Applied Intuition, and open-source users like Roku and Zendesk.
Washington Post Inks OpenAI Licensing Deal for Search
a year ago
- The Washington Post partners with OpenAI to make its content accessible via ChatGPT.
- ChatGPT will display summaries, quotes, and links to Post articles with clear attribution.
- The partnership aims to provide reliable, factual information on complex or fast-moving topics.
- OpenAI has similar deals with over 20 news publishers, including News Corp and Financial Times.
- Some publishers, like The New York Times, have sued OpenAI for copyright infringement.
- Washington Post's global partnerships head emphasizes meeting audiences where they are.
- Over 500 million people use ChatGPT weekly for searches.
- Washington Post has launched AI experiments like 'Ask the Post AI' and 'Climate Answers'.
- The Post uses AI for summaries, audio, and other solutions while remaining LLM-agnostic.
A Review of the Personal Humanoid Robots
a year ago
- Humanoid robots are set to enter homes by 2025, transforming daily life with tasks like chores, tutoring, and elderly care.
- The world is designed for humans, making humanoid robots the most efficient form for automation in human-centric spaces.
- Humanoid robots offer advantages like using existing tools, navigating human environments, and fostering trust through anthropomorphic design.
- Economic benefits include scalability and reprogrammability for multiple roles, reducing the need for specialized robots.
- Psychological factors favor humanoid robots for acceptance in customer-facing and home settings.
- Challenges include complexity, cost, and the 'uncanny valley' effect, but these are expected to diminish with technological advancements.
- A historical timeline shows the evolution from industrial robots to personal robots like the Roomba and Tesla's Optimus.
- The development of humanoid robots is likened to a 'Sputnik moment,' with global competition among nations and companies.
- Security concerns arise as robots connected to the internet could be hijacked for malicious purposes.
- China is aggressively investing in robotics, aiming to dominate the market by 2027, while the U.S. focuses on innovation.
- The global value of humanoid robots is estimated at up to $24 trillion, with significant potential in household and manufacturing applications.
- Technological advancements are reducing costs, making robots more accessible, with prices expected to drop below $10,000 by 2026.
- Leading companies like Tesla, Figure AI, and 1X Technologies are developing robots for various applications, from kitchen assistance to companionship.
- Adoption of robots could address labor shortages, reduce costs of goods and services, and increase leisure time.
- Emerging trends include organized aggression against robots and the rise of a 'Robot Tax' in the European Union.
- The article offers exclusive insights for members on the cultural, philosophical, and practical impacts of humanoid robots.
Exploring on-device AI link previews in Firefox
a year ago
- Firefox Labs 138 introduces an optional experimental feature for link previews to enhance browsing experience.
- Link Preview activation: hover over a link and press Shift + Alt (Option on macOS) to see a card with title, description, image, reading time, and 3 key points.
- The preview card appears in a separate panel, positioned near the cursor, reinforcing it's from Firefox and not the page.
- Initial implementation uses credentialless HTTPS requests to fetch page HTML without executing scripts, sending a custom x-firefox-ai header.
- Metadata like Open Graph tags and Firefox’s Reader View capabilities are used to extract title, description, image, reading time, and key points.
- Key points are generated on-device using wllama (WebAssembly llama.cpp) with SmolLM2-360M from HuggingFace for privacy.
- Performance: first key point appears within 4 seconds, with subsequent points within a second.
- Optimizations include pre-downloading the AI model (369MB) and limiting content provided to the model.
- Future improvements focus on multiple languages, key points quality, and bringing the feature to Android.
- Feedback is encouraged via AI@Mozilla discord #firefox-ai.
Why Open Models Matter
a year ago
- Open-source AI models are faster, more customizable, private, and cost-effective compared to proprietary models.
- Meta's Llama models and other open models have rapidly advanced, rivaling closed systems like GPT-4.
- Open models are shifting AI from being a product to infrastructure, similar to how Linux became foundational.
- Advantages of open models include cost efficiency, total ownership, fine-tuning capabilities, and performance in common tasks.
- Open models enable edge and local AI, improving privacy and reducing dependency on cloud services.
- Transparency in open models enhances security and trust, unlike proprietary models which can drift unpredictably.
- Closed models still hold advantages in compute power and ecosystem integration, but open models are closing the gap.
- Open models democratize AI, allowing broader access and innovation without requiring massive budgets.
- Ethical considerations highlight the importance of open access to prevent monopolization of AI technology.
- The future of AI is leaning towards open models, offering flexibility, transparency, and community-driven innovation.
Hugging Face to sell open-source robots thanks to Pollen Robotics acquisition
a year ago
- Hugging Face acquires Pollen Robotics to expand into open-source robotics.
- Reachy 2, an open-source humanoid robot, is now available for $70,000.
- LeRobot library, launched in 2024, has grown to over 12,000 GitHub stars.
- Pollen Robotics' Reachy 2 features advanced hardware, VR compatibility, and open-source software.
- Hugging Face's fifth acquisition includes notable additions like Gradio and XetHub.
- NVIDIA collaborates with Hugging Face to accelerate LeRobot's AI workflows.
- Reachy 2 is used in institutions like Cornell and Carnegie Mellon for AI research.
- Pollen Robotics has deployed hundreds of robots across 20+ countries.
- Hugging Face aims to democratize robotics with affordable, open solutions.
I won't be vibe coding anymore: a noob's perspective
a year ago
- The author reflects on their early coding experiences, acknowledging the poor quality and slow progress but valuing the personal effort and learning involved.
- They express dissatisfaction with using AI to build apps, feeling it detracts from the learning process and personal growth in coding.
- A realization dawns that coding, like writing, is about the process, problem-solving, and critical thinking, not just the finished product.
- The author decides to return to manual coding, embracing the slow and deliberate process of learning and creating, despite the imperfections.
Teaching LLMs how to solid model
a year ago
- LLMs can create CAD models for simple 3D mechanical parts and are expected to improve rapidly.
- An AI mechanical engineer could handle tasks like material selection, design for manufacturing, and CAD model improvement.
- Training generative models on existing CAD files is a promising approach, with research focusing on diffusion and transformer architectures.
- LLMs can generate solid models using programmatic interfaces like OpenSCAD, leveraging their code generation capabilities.
- OpenSCAD allows users to write scripts that render into solid CAD models, making it suitable for LLM-driven modeling.
- LLMs can utilize general mechanical engineering knowledge, such as correctly sizing holes for bolts.
- An evaluation pipeline was developed to test LLMs' ability to generate CAD models via OpenSCAD, comparing generated STL files to reference models.
- Geometric checks in the evaluation include volume difference, bounding box alignment, chamfer distance, and Hausdorff distance.
- Frontier models like Sonnet 3.5 and Sonnet 3.7 show significant performance improvements in CAD modeling tasks.
- Start-ups like AdamCad and Zoo.dev are developing text-to-CAD products, though initial evaluations show LLMs generating OpenSCAD code outperform Zoo.dev's API.
- Future improvements for LLM-driven CAD modeling include better spatial reasoning, iterative debugging tools, and parameter adjustment UIs.
- GenCAD is expected to become a common tool for mechanical engineers within 6-24 months, with long-term potential to revolutionize CAD suites.

Hasty Briefsbeta

#ai

FurtherAI (YC W24) Is Hiring Software and AI Engineers

Show HN: Dia, an open-weights TTS model for generating realistic dialogue

Spark AI (YC W24) is hiring a full-stack engineer in San Francisco

From 'catch up' to 'catch us': How Google took the lead in enterprise AI

Cekura (Formerly Vocera) (YC F24) Is Hiring

'Immediate red flags': questions raised over 'expert' much quoted in UK press

Columbia student suspended over interview cheating tool raises $5.3M

AI hallucinations lead to a new cyber threat: Slopsquatting

Is SaaS a good business model for drug‑discovery companies?

I Open-Sourced My AI Toy Company That Runs on ESP32 and OpenAI Realtime API

π0.5: A VLA with open-world generalization

More than 20 robots race alongside humans at half marathon in Beijing

Onyx (YC W24) Is Hiring for ML Engineer

Washington Post Inks OpenAI Licensing Deal for Search

A Review of the Personal Humanoid Robots

Exploring on-device AI link previews in Firefox

Why Open Models Matter

Hugging Face to sell open-source robots thanks to Pollen Robotics acquisition

I won't be vibe coding anymore: a noob's perspective

Teaching LLMs how to solid model