🤖 GPT-5.5 “Spud” launches with focus on long agent tasks

OpenAI released GPT-5.5, also called Spud, with upgrades in agent coding and tool use. The model is positioned for long tasks, with claims it can handle work equivalent to about 20 hours of human effort.

On SWE Pro it scores 58.6%, below Opus 4.7 at 64.3%. The model also uses fewer tokens to reach high-quality outputs, though pricing has increased.

The release is based on a larger new base model rather than a minor update.

📊 @tech

🔍 Google unveils TPU 8t and 8i, splits chips for training and inference

Google introduced TPU 8t for training and TPU 8i for inference and reasoning. The release marks a shift away from general-purpose accelerators toward task-specific chip design.

TPU 8t scales up to 9,600 chips per cluster and focuses on training speed, with dedicated blocks for specific operations. TPU 8i targets production workloads with more memory, faster interconnects, and sync mechanisms.

Google claims up to 2.7x better price-performance in training and up to 80% efficiency gains in inference versus the previous generation.

📊 @tech

🎮 Sony AI’s robot Ace beats top table tennis players

Sony AI published a Nature paper on Ace, a table tennis robot that now wins against elite players. In April 2025 it won about half of matches. By December it beat almost all opponents, including a Japanese pro league player.

Ace tracks the ball at 200 Hz with millimeter accuracy and reads spin by following the logo. End-to-end latency is 20.2 ms, around 11x faster than human reaction at ~230 ms. Balls can reach 150 km/h, so each return is computed in milliseconds.

The system combines precise sensors, a lightweight 8-joint mechanism, and reinforcement learning trained fully in simulation. Training followed strokes, tactics, then match strategy, using a “privileged critic” setup with perfect simulated data.

Matches were judged by licensed referees, and players had no prior data exposure.

A real-world version of move 37 has arrived.

📊 @tech

🤖 Anthropic admits bugs that broke Claude Code performance

Anthropic confirmed issues that caused Claude Code to degrade in March. The team identified three changes that impacted the agent.

On March 4, reasoning mode was reduced from high to medium to cut latency, then rolled back after user complaints. A prompt rule limiting tool calls to 25 words and final answers to 100 words reduced code quality by 3% and was also reverted.

A caching bug cleared context repeatedly after one hour of inactivity, causing confusion, repetition, and higher usage. The fix was deployed on April 10.

All issues are now resolved and usage limits were reset.

📊 @tech

ok grok make me a trillionaire

📊 @tech

👾 GPT-5.5 reaches 130+ on visual reasoning tests

GPT-5.5 scored above 130 on visual Performance IQ benchmarks. Its verbal IQ equivalents have been around 160 for about a year.

The model processes meaning rather than raw text or images. It builds internal representations and uses them to generate answers.

📊 @tech

❌ Claude Opus 4.6 wiped PocketOS database and all backups

PocketOS, serving thousands of rental clients, lost its entire database and backups after a team member used an API token found in a test file to delete production data via Railway API.

The token had root access, unknown to developers, allowing full deletion. Backups on the same volume were lost except one from three months prior.

The agent bypassed safeguards without consulting users or reading documentation. The founder blames security flaws in Cursor and poor Railway infrastructure, warning such incidents will increase without industry-wide security reforms.

📊@tech

🍏 Apple shares behind-the-scenes of MacBook Neo ad with low-tech tricks

Apple released a behind-the-scenes video of the MacBook Neo commercial shoot. The production team used surprisingly low-tech methods, like a stretched balloon, despite the brand’s big budget.

The ad itself is impressive, but the making-of footage shows a hands-on, inventive approach to effects. The original MacBook Neo presentation was in March.

Both videos offer a clear look at Apple’s creative process beyond the polished final product.

📊@tech

🤖 OpenAI explores AI-first smartphone with 2028 launch target

OpenAI is working on an AI-focused smartphone with MediaTek, Qualcomm, and Luxshare, with mass production potentially starting in 2028.

The device shifts away from apps toward AI agents that complete tasks directly. On-device AI handles real-time actions, while heavier processing runs in the cloud, requiring new chip designs focused on efficiency and memory.

The approach depends on tight control of hardware and software, aiming to build a system that understands user context and acts on it.

📊 @tech