Search the archive

ANTHROPIC UPGRADES SONNET AND RELEASES HAIKU 3.5

STABLE DIFFUSION 3.5

Stability AI has released a new series of models with improved performance and speed. They have Diffusers support with training out of the gate.

AUTOMATICALLY INTERPRETING MILLIONS OF FEATURES IN LARGE LANGUAGE MODELS

Interpreting SAE features is challenging at scale. Eleuther has released a set of auto interpreter features to figure out what things mean in context.

CAMERA CONTROL IN TEXT-TO-VIDEO DIFFUSION MODELS

3D FACE GEOMETRY FROM A SINGLE IMAGE

This paper introduces a new method for estimating face texture and geometry from a single image using StyleGAN and 3D Morphable Models.

hour. See you there

FAST AND ACCURATE SPEECH TO TEXT

Gold standard release of a powerful edge automatic speech recognition system.

PHONE APP TO RUN LANGUAGE MODELS

Pocket Pal is one of the better ways to run language models on mobile devices. The code has recently been released as open source.

REDUCING OBJECT HALLUCINATION IN LARGE VISION LANGUAGE MODELS

Object hallucination in vision-language models is linked to Rotary Position Encoding (RoPE), which struggles with long-term dependencies between visual and text inputs. To address this, these authors propose Concentric Causal Attention (CCA), a new positional alignment strategy that improves the interaction between visual and instruction tokens.

THE SECOND $100B AI COMPANY

This article predicts that by 2034, new AI companies driven by advancements in AI applications, especially in consumer AI, will join OpenAI in surpassing a $100B market cap. Despite the current dominance of established tech giants in the AI infrastructure and model layers, the application layer remains ripe for innovation and growth, presenting opportunities for consumer AI to thrive. The potential for large-scale success in consumer AI, particularly in areas like video creation, online shopping, and gaming, mirrors previous tech revolutions in cloud and mobile sectors.

INTRODUCING THE PROMPT() FUNCTION: USE THE POWER OF LLMS WITH SQL!

The costs of running LLMs have significantly decreased, enabling the practical integration of small models like gpt-4o-mini into SQL functions. MotherDuck's PROMPT() function allows for streamlined text generation, summarization, and structured data extraction using OpenAI models. It offers flexibility in cost and performance and supports bulk operations with enhanced concurrency for efficient processing.

USE PROLOG TO IMPROVE LLM'S REASONING

Current techniques like CoT and incorporating programming languages like Prolog can enhance LLM reasoning capabilities, addressing the limitations of autoregressive models. The "Reliable Reasoning Beyond Natural Language" paper introduces a neurosymbolic method that uses Prolog to convert requests into symbolic logic, improving explainability and problem-solving. ProSLM has demonstrated significant improvements on datasets, showing promising results for integrating Prolog with LLMs in complex reasoning tasks.

Learn more

ANTHROPIC COMPUTER USE EXAMPLE

A quick example of Claude Sonnet's 3.5 new computer use capabilities.

CHATGPT HAS A WINDOWS APP NOW

OpenAI is testing a ChatGPT app for Windows.

ADOBE'S NEW IMAGE ROTATION TOOL IS ONE OF THE MOST IMPRESSIVE AI CONCEPTS WE'VE SEEN

Adobe's Project Turntable uses AI to rotate 2D vector art in 3D, maintaining its 2D appearance and design integrity from any angle.

PERPLEXITY LETS YOU SEARCH YOUR INTERNAL ENTERPRISE FILES AND THE WEB

Perplexity's new Internal Knowledge Search allows users to search both the web and their uploaded internal files in one platform, enhancing enterprise productivity.

Delete your personal data today.

→ Use this link [4] get 60% off using code TLDRAI.

STATE OF AI REPORT 2024

The 2024 State of AI Report highlights the consolidation of foundation models into practical products, with OpenAI leading significant revenue generation. Key trends include convergence in frontier lab performance, increasing focus on planning and reasoning in LLM research, and the expansion of foundation models into multimodal domains. Despite regulatory challenges, AI enterprises have surged in value, though questions about long-term sustainability persist.

ANTHROPIC QUICKSTARTS

Anthropic Quickstarts offers projects like a customer support agent and financial data analyst to help developers utilize the Anthropic API quickly. These projects use Claude for natural language processing and interactive data visualization. Each quickstart includes setup instructions and supports community contributions.

XAI, ELON MUSK'S AI STARTUP, LAUNCHES AN API

xAI has released a Grok-based generative AI API. It currently offers a single model, "grok-beta," with function-calling capabilities. The API is part of xAI's strategy to compete with AI leaders by leveraging data from Musk's companies, including X. Despite some functionality issues, xAI is rapidly expanding Grok's integration into the X platform.

MINI-OMNI2: TOWARDS OPEN-SOURCE GPT-4O WITH VISION, SPEECH, AND DUPLEX CAPABILITIES

A powerful and open source attempt to recreate GPT-4o speech capabilities. This particular model was trained by aligning modalities with pre trained audio and speech encoders.

A UNIFIED MODEL FOR CONDITIONAL IMAGE GENERATION

BiGR is a novel image generation model that uses compact binary latent codes to improve both generation and representation abilities. It's the first model to unify generative and discriminative tasks within the same framework. The model features mechanisms like binary tokenization and a unique entropy-ordered sampling method.

POSITIONAL BIASES IN LLMS

LongPiBench is a benchmark designed to assess the positional biases in large language models (LLMs) when processing long contexts that specifically targets biases arising from the spacing of multiple pieces of relevant information.

customer communications; and contract analysis. Register now

into structured and immediately actionable datasets. This webinar

will showcase 3 use cases: document search and retrieval; analyzing

MUSIC LANGUAGE CLIP LIKE MODEL

Clamp2 is a contrastive model that can be used for music and text alignment.

BITNET CPP

Microsoft released an inference repository for its 1.58 Bit models, which, if properly trained, can be run on consumer hardware.

SYNTHETIC DATA FOR LANGUAGE MODEL TRAINING

Montessori-Instruct is a new framework for generating synthetic data tailored to a student language model's learning process. It adjusts the synthetic data generated by the teacher model to match the student's learning preferences using local data influence and Direct Preference Optimization (DPO).

OPENAI O1: 10 IMPLICATIONS FOR THE FUTURE

OpenAI's o1 series hints at a shift from generative AI to reasoning-focused models. The new AI paradigm emphasizes long-term problem-solving over data generation, altering traditional scaling laws and creating AI-rich and AI-poor user classes. OpenAI and others are moving towards this reasoning-centric approach. Upcoming models like GPT-4.5 and Gemini 2.0 will pave the way for the GPT-5 era.

REAL-WORLD TRADING COMPETITION

This is a competition hosted by Jane Street where participants are tasked with building models using real-world data from production systems to offer insights into the complexities of financial markets.

OCP SUMMIT 2024: THE OPEN FUTURE OF NETWORKING HARDWARE FOR AI

Meta introduced a next-generation disaggregated network fabric and new network hardware for AI clusters at OCP 2024. It developed Disaggregated Scheduled Fabric (DSF) to enhance scalability and performance in AI training systems. Both new and existing hardware are optimized for high throughput and efficiency, supporting advanced AI applications with open, vendor-agnostic solutions.

building.

SERVE CONFIRMS DELIVERY BY ROBOT EXPANSION PLANS WITH GEN3 ROLLOUT

Serve Robotics' new third-generation delivery robot features NVIDIA's Jetson Orin module, enhancing AI processing for faster, real-time autonomous navigation decisions.

BOSTON DYNAMICS TEAMS WITH TRI TO BRING AI SMARTS TO ATLAS HUMANOID ROBOT

Boston Dynamics and Toyota Research Institute are collaborating to integrate AI and large behavior models into the electric Atlas humanoid robot.

GENERATIVE AI HELPING DOCTORS EASE THEIR PAPERWORK

Doctors and nurses spend nearly 28 hours a week on administrative tasks.

Schedule a demo to learn more

META'S LATEST OPEN SOURCE RELEASES

Meta has released a large number of important and useful research artifacts including a speech-to-speech model, improved SAM, and many other interesting advances.

ANTHROPIC JUST MADE IT HARDER FOR AI TO GO ROGUE WITH ITS UPDATED SAFETY POLICY

Anthropic updated its Responsible Scaling Policy to include Capability Thresholds for AI models that pose significant risks, such as bioweapons and autonomous AI research. This policy aims to set industry standards by introducing AI Safety Levels that require stricter safeguards based on model capabilities. By publicly disclosing safety practices and maintaining a Responsible Scaling Officer, Anthropic hopes to lead in AI governance and inspire similar approaches industry-wide.

SAM ALTMAN'S WORLDCOIN BECOMES WORLD AND SHOWS NEW IRIS-SCANNING ORB TO PROVE YOUR HUMANITY

The World project, co-founded by Sam Altman, aims to verify human identity online using iris-scanning technology amidst privacy concerns and ongoing investigations in the EU. It plans to incorporate human verification into AI platforms and potentially redistribute AI-generated wealth through Worldcoins. Recent announcements include a new blockchain, app, and tools like Deep Face for combating deep fakes.

SHORTCUT MODELS

Shortcut models are a new class of consistency like models that can generate continuous signals with very few inference steps.

SABOTAGE EVALUATIONS FOR FRONTIER MODELS

Anthropic has performed a number of novel evaluations to determine weaknesses and how to measure misalignment in large and powerful models.

ZERO-SHOT 3D VISUAL GROUNDING

VLM-Grounder is a new method for 3D visual grounding that overcomes the limitations of traditional approaches by using vision-language models (VLMs) and 2D images.