Saturday — August 2, 2025

Atlassian terminates 150 staff to replace them with AI, Cerebras launches Code Pro and Code Max plans for fast code completions, and researchers develop SpeedLLM, an FPGA co-design for large language model inference acceleration.

News

Cerebras Code

Cerebras is launching two new plans, Code Pro ($50/month) and Code Max ($200/month), which provide access to the Qwen3-Coder AI model, offering fast and high-context code completions with speeds of up to 2,000 tokens per second. These plans allow users to integrate the model with their own IDE, with Code Pro suitable for indie devs and small projects, and Code Max ideal for full-time development and heavy coding workflows.

Atlassian terminates 150 staff

Atlassian, a major Australian tech company, has terminated 150 staff via a pre-recorded video message from CEO Mike Cannon-Brookes, with those affected to be largely replaced by AI-embedded customer contact solutions. The terminated staff, reportedly from the company's European operations, will receive six months' pay, and the move is part of Atlassian's push to increase its use of artificial intelligence in customer service.

GPT-5 is already (ostensibly) available via API

Your request has been blocked due to a network policy, and you're advised to log in or create an account to continue browsing. If you're running a script or application, you should register or sign in with your developer credentials and ensure your User-Agent is unique and descriptive to resolve the issue.

LLM leaderboard – Comparing models from OpenAI, Google, DeepSeek and others

Artificial Analysis has compiled a leaderboard comparing over 100 AI models from various providers, including OpenAI, Google, and DeepSeek, across metrics such as intelligence, price, output speed, and latency. The top models in each category include Grok 4 and o3-pro for intelligence, Gemini 2.5 Flash-Lite for output speed, Aya Expanse 8B for latency, and Gemma 3n E4B for price.

The AI age is the "age of no consent"

The author argues that we are living in an "age of no consent" where AI companies and tech giants are prioritizing their own interests over user needs and privacy, often using deceptive tactics to force users into adopting AI-powered features. This shift has led to a dilution of the concept of "ethical" and "empathy" in tech, with companies exploiting user data without consent and ignoring user feedback, ultimately reversing the traditional relationship between companies and their customers.

Research

SpeedLLM: An FPGA Co-Design of Large Language Model Inference Accelerator

SpeedLLM, a neural network accelerator, is designed to optimize edge computing performance on the Xilinx Alevo U280 platform for the Tinyllama framework, utilizing innovations such as data stream parallelism and operator fusion. The SpeedLLM architecture achieves significant improvements, outperforming traditional implementations with up to 4.8 times faster performance and 1.18 times lower energy consumption.

TinyTroupe: An LLM-Powered Multiagent Persona Simulation Toolkit (OSS Paper)

Recent advances in Large Language Models have led to the development of LLM-powered Multiagent Systems, but existing tools lack capabilities for realistic human behavior simulation, prompting the creation of TinyTroupe, a simulation toolkit. TinyTroupe enables detailed persona definitions and programmatic control via LLM-driven mechanisms, allowing for the formulation and solution of behavioral problems, and is available as an open-source Python implementation.

Emotion Detection in Older Adults Using Signals from Wearable Sensors

Researchers investigated an edge-based approach to emotion identification in older adults using physiological signals from wearable sensors, achieving a high accuracy rate with a 0.782 r2 score. This non-obtrusive method has significant implications for individuals with cognitive impairments, such as Alzheimer's and PTSD, and paves the way for privacy-preserving emotion recognition systems in real-world settings.

Interpretable EEG-to-Image Generation with Semantic Prompts

Researchers have developed a model that can decode visual experiences from brain signals using EEG, bypassing direct image generation by aligning EEG signals with semantic captions generated by a language model. This approach yields state-of-the-art visual decoding results and provides interpretable insights into neurocognitive pathways, demonstrating the effectiveness of structured semantic mediation in visual decoding from EEG signals.

Cot-Self-Instruct: Synthetic prompts for reasoning and non-reasoning tasks

CoT-Self-Instruct is a synthetic data generation method that uses Chain-of-Thought to generate high-quality training data for large language models, which outperforms existing datasets in verifiable reasoning tasks. The method also surpasses human or standard self-instruct prompts in non-verifiable instruction-following tasks, achieving better performance on various evaluation benchmarks.

Code

Show HN: TraceRoot – Open-source agentic debugging for distributed services

TraceRoot is an open-source debugging platform that uses AI-powered analysis to help engineers fix production issues 10 times faster by combining structured traces, logs, and source code context. The platform is designed with principles such as intelligence, real-time tracing and logging, structured information, integration, and developer-friendly interface, making it a powerful tool for debugging and tracing applications.

Eigent, a Multi-agent Workforce desktop application

Eigent is a multi-agent workforce desktop application that enables users to build, manage, and deploy a custom AI workforce to automate complex workflows. It offers various features, including parallel execution, customization, and privacy protection, and is available in cloud, self-hosting, and enterprise versions, with a 100% open-source codebase.

Remove AI Summaries

The AI summary remover is a Chromium extension that removes AI-generated summaries from news articles on select Norwegian websites, including NRK, VG, and Dagbladet. The extension is currently available for manual installation from source, with plans to publish it on the Chrome webstore, and was motivated by concerns about the impact of AI on human intelligence and education.

Show HN: Agentic AI Frameworks on AWS (LangGraph,Strands,CrewAI,Arize,Mem0)

This repository provides examples and reference architectures for building autonomous agents using popular frameworks on AWS services, covering various layers of the agent stack and industry verticals. The repository includes a range of examples, AWS blogs, and workshops that demonstrate how to leverage AWS services to create production-ready agent applications, and welcomes contributions from the community.

Show HN: Inworld TTS open sourcing and technical report

This repository contains the training and modeling code for Inworld TTS-1 and TTS-1-Max models, allowing users to pre-train, fine-tune, or align with RL their arbitrary SpeechLM-based TTS model. The code provides features such as distributed training, data pipeline, and modeling with SpeechLM and 1D audio-codecs, and includes a quick start guide and example usage for training and inference.

    Atlassian terminates 150 staff to replace them with AI, Cerebras launches Code Pro and Code Max plans for fast code completions, and researchers develop SpeedLLM, an FPGA co-design for large language model inference acceleration.