Search the archive

posture. Get the guide from BigID

HOW TO CREATE HIGH-QUALITY SYNTHETIC DATA FOR FINE-TUNING LLMS

Gretel released new data comparing the performance of synthetic data to human expert-curated datasets.

THE ALL-SEEING AI WEBCAM

Artist Dries Depoorter's AI experiments, including a Kylie Jenner voice clone offering selfie advice, highlight his provocative art that toys with surveillance and privacy using unsecured webcam footage and AI technology.

MRSCRAPER (PRODUCT LAUNCH)

Get and scrape any data at scale and get past captcha blockers.

LLAMA 3.1 AND 405B MODELS

Meta has released its first open frontier AI model along with training details. The model is on par with top closed models with a 128k context length, chat models, and a great open system.

HARVEY RAISES $100M SERIES C

The legal startup, now a unicorn, has raised from Google ventures and others to continue its push into big law firms.

OPEN SOURCE IS THE PATH FORWARD

Mark Zuckerberg outlined why open source is key to Meta's strategy and how Meta plans to support this work.

IMPROVED TEST-TIME ADAPTATION

STAble Memory rePlay (STAMP) is a method that addresses distribution shifts between training and test data using only unlabeled test data. Unlike existing methods, STAMP excels in both recognizing known classes and rejecting outliers during inference.

VIDEO POINT TRACKING

LocoTrack is an advanced model for tracking any point across video sequences. It uses local 4D correlation and bidirectional correspondence for precise tracking. LocoTrack operates nearly six times faster than current leading models.

STAR OPERATION TO TRAIN NEURAL NETWORKS

The star operation is a method that enriches AI models by projecting data into complex, high-dimensional spaces without the need for bulky architectures.

as Python data objects. Star the repo

LLAMA AGENT STACK

Meta has released an example system for using its Llama models as agents to perform a variety of tasks.

TEXT-DRIVEN STYLIZATION WITH DIFFUSION MODELS

Artist is a training-free method that controls content and style generation in pretrained diffusion models for text-driven stylization.

OPEN-WORLD EXPLORATION IN MINECRAFT

Odyssey is a new framework that equips large language model-based agents with advanced skills for exploring Minecraft.

GUMLOOP SEED ROUND

Gumloop raised a $3.1M Seed round led by First Round Capital with participation from YC and co-founders at Instacart, Dropbox, and Airtable. Gumloop is a no-code AI automation platform that lets any employee within a business build their own AI tooling and have as much impact as an engineer.

AI DEVELOPMENT KITS: TENSTORRENT UPDATE

Tenstorrent has launched affordable AI development hardware, including the Wormhole n150 and n300 PCIe cards, priced at $999 and $1,399, as well as developer workstations, the air-cooled TT-LoudBox for $12,000 and water-cooled TT-QuietBox for $15,000, both designed to aid AI development with a focus on connectivity and scaled-out performance.

AI PREDICTS DROUGHTS A YEAR IN ADVANCE

Skoltech and Sber researchers have developed AI models that can predict droughts several months to a year in advance, improving risk management for agriculture, insurance, and finance sectors. The models, tested across various climates, utilize spatiotemporal neural networks and publicly available data. Russia's largest bank plans to integrate these findings into its risk assessment systems.

Watch the Confluent webinar now

WHAT DOES MONEY LOOK LIKE IN AN AI UTOPIA?

It's unclear what role money will play in a future AI utopia.

SAMSUNG IS POURING RESEARCH INTO “AI PHONES”

THIS IS HOW MUCH DATA DOES AI CREATES EVERY MINUTE

Every sixty seconds, nearly $300K is spent on AI, 52 college papers are plagiarized using AI, and nearly 20K images are created using text-to-image algorithms.

COHERE SEES VALUATION SOAR TO $5.5B AFTER NEW FUNDING ROUND

Cohere, a Canadian AI startup focused on large language models, has reached a $5.5B valuation following a $500M Series D funding round. The new funds aim to enhance its enterprise-grade AI tech for global business efficiency. Key investors include PSP Investments, Cisco, Fujitsu, AMD Ventures, and EDC.

FIGMA AI UPDATE

Figma briefly rolled back its 'Make Designs' AI feature from its limited beta after realizing it generated UI designs that too closely resembled existing apps. The feature uses off-the-shelf AI models like GPT-4 and Amazon's Titan and it needs refinement to ensure originality. Figma aims to re-enable the feature with improved QA processes to better assist designers in leveraging AI for efficient design creation.

ELEVENLABS TURBO 2.5 MODEL

ElevenLabs has introduced a new model, Turbo 2.5, that unlocks high-quality low-latency conversational AI for nearly 80% of the world - Hindi, French, Spanish, Mandarin, and 27 other languages. For the first time, it supports Vietnamese, Hungarian, and Norwegian text-to-speech. English is now 25% faster compared to Turbo v2.

ASSISTANT BENCHMARK

A new benchmark to assist in determining a model's agent-like abilities.

RANKING-BASED LOSSES FOR OBJECT DETECTION

Bucketed Ranking-based Losses improve the efficiency of ranking-based loss functions in object detection.

INTERPRETABILITY SURVIVAL ANALYSIS

SurvReLU is a deep survival model that uses rectified linear unit (ReLU) networks, bridging the gap between "black-box" neural networks and "white-box" tree-based models.

ORCH

Orch is a library for building language model powered applications and agents in the Rust programming language.

ADVANCING TEXT-TO-3D GENERATION

PlacidDreamer is a framework for text-to-3D generation that resolves issues with previous methods by unifying generation directions and addressing over-saturation.

HEAD POSE ESTIMATION

Researchers developed the head Translation, Rotation, and face Geometry network (TRG) to improve head pose estimation, particularly focusing on head translations.

USING LLMS FOR EVALUATION

As large language models become more capable, evaluating them accurately remains challenging due to their diverse tasks. While human evaluation is the gold standard, it's costly and slow. Using LLMs themselves as evaluators offers a scalable, cost-effective alternative, despite potential biases such as positional and verbosity bias, which can be mitigated through techniques like randomizing output positions and using multiple evidence calibrations.

THREE ARCHETYPES OF AI APPLICATION STARTUPS

AI applications are evolving into three common archetypes: AI Copilots that assist with tasks, AI Colleagues that autonomously manage specific tasks alongside human workers, and AI-Native Services which offer end-to-end services that integrate AI with human input. GitHub Copilot and Devin exemplify AI Copilots and Colleagues, aiding in coding and engineering, respectively. AI-Native Services, such as Pilot for bookkeeping, compete with traditional service firms by offering automated solutions in areas like law and accounting.

INSIDE THE FIGHT OVER CALIFORNIA'S NEW AI BILL

California state Senator Scott Wiener's Safe and Secure Innovation for Frontier Artificial Intelligence Models bill requires companies training 'frontier models' that cost more than $100 million to do safety testing and be able to shut off their models in the event of a safety incident. The bill has received heavy criticism from the tech industry. It will affect everyone who does business in California, not just companies that develop their models in California. This article contains an interview with Wiener about the bill and its critics.

SHADOW (PRODUCT LAUNCH)

Shadow is a botless AI meeting assistant for automated productivity, from note-taking to action item completions.

HOW FAST CAN STRUCTURED GRAMMAR GENERATION BE

Structured generation in language models is quickly being tackled by the open source community.

HEYGEN INTERACTIVE AVATAR (PRODUCT LAUNCH)

HeyGen Interactive Avatar is a lifelike avatar that can communicate in real time with your customers, providing support and sharing knowledge effortlessly.

Use code TLDRAI today for a 60% discount →

ARTIFICIAL AGENCY RAISES $16M TO USE AI TO MAKE NPCS FEEL MORE REALISTIC IN VIDEO GAMES

Former Google DeepMind scientists at Artificial Agency unveiled an AI behavior engine for dynamic NPCs in video games. They have raised $16 million to enhance in-game interactions. They're working with renowned AAA studios and expecting widespread adoption by 2025 despite potential cost implications. Their engine offers more realistic and responsive game characters without predefined scripts.

INSIDE THE UN'S AI POLICY GRAB

The United Nations aims to create a centralized AI forum to address the fragmented global efforts in AI governance.

NVIDIA PREPARING VERSION OF NEW FLAGSHIP AI CHIP FOR CHINESE MARKET

Nvidia is developing a new AI chip, the B20, tailored to comply with U.S. export controls for the Chinese market, leveraging its partnership with distributor Inspur. Its advanced H20 chip has reportedly seen a rapid growth in sales in China, with projections of selling over 1 million units worth $12 billion this year. U.S. pressure on semiconductor exports continues, with possible further restrictions and control measures on AI model development.

APPLE DCLM (HUGGING FACE HUB)

Apple has released everything, including code and data, for its 7B parameter model, which is competitive with Mistral.

SCALING LAWS FOR VOCABULARY SIZE

Larger models need a larger vocab, according to this new research which looks at scaling laws for vocabulary size.

ENHANCED MULTIMODAL LANGUAGE MODELS WITH MOME

Researchers propose the Mixture of Multimodal Experts (MoME) to tackle the issue of task interference in generalist Multimodal Large Language Models (MLLMs).

SCIENTIFIC COMPUTING IN MOJO

Endia is an array based programming library for Mojo designed to assist in various scientific and machine learning tasks.