New "memory sources" feature shows which saved memories influenced each response; personalization from past chats, files, and Gmail rolling out to Plus/Pro (web) first

Until now, ChatGPT users could view a list of saved memories but had no way to trace which ones affected a specific reply. Over the past year, OpenAI broadened personalization by letting the model draw on past conversations and uploaded files. This update adds a "memory sources" panel that surfaces the exact memories behind each answer, making the personalization layer transparent for the first time.

Users can now audit whether unintended personal details are shaping responses — a meaningful privacy control for everyday use. On the flip side, intentionally building a well-structured memory bank can eliminate the need to re-explain context each session. Since the rollout starts with Plus and Pro (web) users, this feature is worth factoring in when comparing subscription tiers.

Google Home updated to Gemini 3.1 — complex multi-step smart home tasks can now be triggered with a single natural-language command

Smart home assistants have traditionally operated on a one-command-one-action basis — turn on the lights, set the thermostat. Google Home included this limitation, requiring users to set up manual routines or issue multiple commands. The Gemini 3.1 update closes that gap by enabling the assistant to parse and execute multi-step sequences from a single conversational sentence.

Scenarios like an evening movie setup or morning departure routine — coordinating lights, temperature, and blinds simultaneously — now require just one sentence. This lowers the barrier for users who found smart home routines too tedious to configure. For households deep in the Google ecosystem, the tighter link with Calendar, Search, and other services raises the practical ceiling of voice control.

ARR exceeded $500M (approx. 43% YoY growth), valuation at $11B, Series D total over $500M with BlackRock, Nvidia, Jamie Foxx and others

ElevenLabs built its reputation on expressive text-to-speech and high-fidelity voice cloning across 32+ languages, raising progressively larger rounds. Over the past two years, use cases expanded from content creation into enterprise applications — video dubbing, call center agents, and real-time conversational AI. ARR had been accelerating rapidly as these production workloads scaled.

Revenue of $500M+ growing at 43% annually signals voice AI has crossed from experimentation into core enterprise infrastructure. Nvidia's participation suggests tighter GPU-level integration for high-volume real-time audio workloads. For developers evaluating voice APIs, this round substantially reduces counterparty risk and signals a long runway for ElevenLabs as a platform investment.

Gemma 4 MTP (Multi-Token Prediction) released — delivers 2x+ inference speedup for coding tasks on Mac via speculative decoding

Gemma is Google's family of open-weight models designed to run efficiently on local hardware. Like most language models, earlier Gemma versions generated one token at a time, setting a ceiling on response speed regardless of hardware. The new MTP variant predicts several tokens simultaneously, effectively cutting the number of forward passes per sentence.

Developers running local AI on Mac get a meaningful speed boost for coding tasks without cloud APIs or heavier hardware. For latency-sensitive tasks — inline completions, quick script generation, iterative debugging — this makes local open-weight inference more competitive with hosted APIs. Teams evaluating self-hosted or on-premises models now have a faster Gemma option to benchmark.

Prompt injection via Morse code tricked a downstream crypto bot into transferring $200K — highlights risk of executing LLM output without validation

AI agents that connect language models to external tools — APIs, wallets, file systems — have been growing rapidly in production. Security researchers have long warned about prompt injection, where malicious input tricks a model into executing unintended instructions. Real-world financial losses from this vector were rare, with most demonstrations confined to research environments.

Any pipeline that passes AI output directly to an action system is vulnerable if the model can be fed adversarial input through any channel, including encoded formats like Morse. Developers building agentic systems need a validation layer between the model's output and any high-stakes action, regardless of model capability. This incident makes a concrete case for human-in-the-loop confirmation on irreversible operations.