OpenAI launches GPT-4.5 for creative tasks, Sesame and Inception Labs release new models, and DeepSeek introduces AI tools
New Models and Releases
OpenAI Introduces GPT-4.5: GPT-4.5, a new AI model from OpenAI, has been introduced with a focus on creative tasks and agentic planning. It features a 128k context length and is currently available in a research preview. The model is noted for its high cost and significant reduction in hallucinations, improving accuracy and reliability. Pricing is set at $75.00 per 1M tokens for input, $37.50 per 1M tokens for cached input, and $150.00 per 1M tokens for output.
- Introducing GPT-4.5
- Watch the ChatGPT 4.5 Reveal
- ChatGPT 4.5 is the #2 best coder in the world on LiveBench
- Sam Altman: GPT-4.5 is a giant expensive model
- GPT-4.5 crushed my personal benchmark
- Empirical evidence that GPT-4.5 is beating scaling expectations
- I feel like some people are missing the point of GPT-4.5
- GPT-4.5's Low Hallucination Rate is a Game-Changer
- Introduction to GPT-4.5 discussion
- GPT-4.5 has an API price of $75/1M input and $150/1M output
- Thoughts on GPT-4.5 and why it's important
- GPT 4.5 released, here's benchmarks
Sesame Develops Conversational Voice Model: Sesame has created a new conversational voice model that rivals OpenAI's Advanced Voice Mode. The model is capable of maintaining context and engaging in realistic conversations but has limitations such as not detecting emotion or sarcasm. Sesame plans to release their models under an Apache 2.0 license in the future.
Inception Labs Releases Diffusion-Based Coding LLM: Inception Labs has introduced a new diffusion-based coding LLM that is 10x faster in token generation than transformer-based LLMs, achieving approximately 1000 tokens per second on the H100 hardware. The model generates all tokens at once and then refines them.
- A diffusion based 'small' coding LLM that is 10x faster in token generation than transformer based LLMs
- 2 diffusion LLMs in one day
DeepSeek Releases 3FS and smallpond: DeepSeek has introduced 3FS, a high-performance distributed file system designed for AI training and inference workloads, leveraging modern SSDs and RDMA networks. Additionally, they released smallpond, a lightweight data processing framework built on DuckDB and 3FS.
Services and Platforms
Novel Forge AI-Powered Writing Tool: Novel Forge, a new AI-powered writing tool, allows users to interact with AI chatbots within their writing projects, edit AI responses, and integrate with various AI models and platforms. The tool is designed for local use and is available for Windows.
Research and Benchmarks
Microsoft Introduces LongRoPE2: Microsoft researchers have introduced LongRoPE2, a novel approach that extends the effective context window of pre-trained large language models (LLMs) to 128K tokens while preserving short-context performance. The code for LongRoPE2 is available on GitHub.
Regional Availability
OpenAI Expands Sora Availability: OpenAI has announced the availability of Sora to Plus and Pro users in the EU, the UK, Switzerland, Norway, Liechtenstein, and Iceland. This expansion follows the initial release and includes regions where access can be verified or circumvented using a VPN.
User-Generated Models
French Reasoning Model: A user fine-tuned a 7B LLM based on Qwen 2.5 to improve its reasoning abilities in French, achieving performance comparable to R1 Distil 7B on math benchmarks with minimal knowledge degradation. The training cost was only $20, and the model and data are available on Hugging Face.