# Feb 27, 2026

## Experts sound alarm after ChatGPT Health fails to recognise medical emergencies

- **ChatGPT Health** failed to recognize medical emergencies, advising patients to wait for appointments in **51.6%** of urgent cases, raising concerns about potential harm and fatalities.

## Nano Banana 2: Google's latest AI image generation model

- **Nano Banana 2** merges the advanced capabilities of **Nano Banana Pro** with the **speed of Gemini Flash**, enabling rapid image generation and editing across Google products.

## OpenAI raises $110B on $730B pre-money valuation

- **OpenAI has secured $110 billion** in private funding, marking a historic investment round led by **$50 billion from Amazon** and **$30 billion each from Nvidia and SoftBank**, against a **$730 billion pre-money valuation**.

## I rendered 1,418 confusables over 230 fonts. Most aren't confusable to the eye

- **96.5% of Unicode confusable pairs score low on visual similarity**, indicating that most pairs are not easily confused visually, yet **82 pairs are pixel-identical** in at least one font, posing a potential security risk.

## Can you reverse engineer our neural network?

- **Jane Street challenges enthusiasts** to reverse engineer their neural network, inviting a deeper understanding of its architecture and functionality.

## We gave terabytes of CI logs to an LLM

- **LLMs excel at SQL** by autonomously generating queries to analyze **1.5 billion CI log lines** weekly, enabling rapid identification of issues, such as tracing flaky tests back to dependency changes in mere seconds.

## Parakeet.cpp – Parakeet ASR inference in pure C++ with Metal GPU acceleration

- **Fast speech recognition** is achieved with NVIDIA's **Parakeet** models in pure C++, boasting **~27ms encoder inference** on Apple Silicon GPUs, which is **96x faster than CPU** for 10s audio using the 110M model.

## Lidar waveforms are worth 40x128x33 words

- **Lidar's potential is enhanced** through a learned DSP that processes full waveforms with a transformer architecture, improving point cloud fidelity by leveraging adjacent waveform features.

## Implementing Better Pytorch Schedulers

- **Enhanced flexibility**: The new scheduling suite for PyTorch allows for dynamic adjustments of **any optimizer hyperparameter** (e.g., LR, momentum, betas) through user-defined functions, addressing the limitations of existing schedulers that only modify learning rates.

## PerpetualBooster v1.9.0 - GBM with no hyperparameter tuning, now with built-in causal ML, drift detection, and conformal prediction

- **PerpetualBooster v1.9.0** introduces **built-in causal ML**, **drift detection**, and **conformal prediction**, enhancing its gradient boosting capabilities without hyperparameter tuning, using a single `budget` parameter for model fitting.

## Evaluating the inference efficiency of Sparse+Linear Hybrid Architectures (MiniCPM-SALA)

- **MiniCPM-SALA** (Sparse+Linear) is being benchmarked in the **SOAR 2026** performance sprint by OpenBMB and NVIDIA, focusing on optimizing **sparse operator fusion** and **KV-cache efficiency** for ultra-long context.

## ASURA: Recursive LMs done right

- **Recursive LMs (RLMs)** can achieve significant performance improvements over traditional models by employing simple techniques, challenging the notion that they are ineffective in language tasks.

## Vibe coded Lovable-hosted app littered with basic flaws exposed 18K users

- **Lovable's platform exposed over 18,000 user records** due to critical vulnerabilities in an app, highlighting the risks of relying on AI-generated code without proper security measures.

## OpenAI and Amazon announce strategic partnership

## Introducing the Stateful Runtime Environment for Agents in Amazon Bedrock