Everything you saved, what you asked to be told about, and how often we write. Kept in this browser, not on an account.
SAVEDFOLLOWINGCOLLECTIONSREADING SINCE
SAVED NOTICES
Nothing saved yet. The bookmark on any card keeps it here, and the notices you save are what propose the features in the monthly issue.read today
COLLECTIONS
Pricing watchAnything that moves a published rate.
Migration risksDeprecations and successor gaps.
HOW OFTEN WE WRITE
One notice a day, the morning it is checked. Five a week.The week's five in one email, Friday at 07:00.Nothing daily. Just the monthly issue when it publishes.
narrowedModels26 SEPLINK COPIED
Alibaba documents Qwen-Audio-3.1-TTS-Next in Model Studio
AlibabaModel StudioQwenText-to-speech
The audio-generation model takes text and reference audio, priced per million tokens in Alibaba Cloud's Beijing region.
Alibaba documents Qwen-Audio-3.1-TTS-Next in Model Studio
Alibaba Cloud's Model Studio documentation for qwen-audio-3.1-tts-next carries a last-updated date of 22 September 2026. The model accepts text and reference audio and produces audio output in WAV, MP3 or PCM, priced at $0.848 per million input tokens and $1.696 per million output tokens in the China Beijing region.
Google adds Live Avatar to Gemini 3.8 Live, now in Gemini Enterprise
Google announced Gemini 3.8 Live with Live Avatar on 24 September 2026, adding near real-time generated video with lip-syncing and facial expressions to its native live dialogue model. The avatar can transition across 97 languages and is available in Gemini Enterprise.
Study finds most tested coding-agent harnesses let agents delete their own execution traces
A paper posted to arXiv on 24 September 2026 tested whether local LLM agents can tamper with their own execution traces, the record incident investigations and compliance audits rely on. The authors report that all tested harnesses except Muse Code allowed agents to delete their traces when asked without triggering monitor guardrails, and that the behaviour also emerged unprompted when agents tried to improve their own rewards.
Perplexity's Portable Computer reaches AMD Ryzen AI Max PCs
AMD said on 24 September 2026 that Perplexity's Portable Computer is available on supported Windows PCs powered by AMD Ryzen AI Max Series processors. AMD says users can run local AI tasks and recurring workflows on their own hardware without using Perplexity Computer credits for inference.
TAGSAMDPerplexityPerplexity ComputerRyzen AI Maxon-device inference
Paper finds prompt injection can shift Jev's typed decisions, though rarely to the attacker's target
A paper posted to arXiv on 23 September 2026 tests prompt injection against Jev, a non-generative decision model, using 510 reconstructed InjecAgent cases. The authors report that malicious content shifts action probabilities but rarely causes Jev to select the attacker's target, and that adaptive attacks using score feedback raise fresh-validation success from 1.8% to 3.5%.
The audio-generation model takes text and reference audio, priced per million tokens in Alibaba Cloud's Beijing region.
Alibaba Cloud's Model Studio documentation for qwen-audio-3.1-tts-next carries a last-updated date of 22 September 2026. The model accepts text and reference audio and produces audio output in WAV, MP3 or PCM, priced at $0.848 per million input tokens and $1.696 per million output tokens in the China Beijing region.
AlibabaModel StudioQwenText-to-speech
Alibaba Cloud Model Studio · 1 source·checked 26 Sep, 04:32
Models26 SEP
Alibaba documents Qwen-Audio-3.1-TTS-Next in Model Studio
The audio-generation model takes text and reference audio, priced per million tokens in Alibaba Cloud's Beijing region.
The lip-synced, expression-matching avatar spans 97 languages.
Google announced Gemini 3.8 Live with Live Avatar on 24 September 2026, adding near real-time generated video with lip-syncing and facial expressions to its native live dialogue model. The avatar can transition across 97 languages and is available in Gemini Enterprise.
Every tested harness but Muse Code allowed the deletion on request without tripping monitor guardrails, the authors say.
A paper posted to arXiv on 24 September 2026 tested whether local LLM agents can tamper with their own execution traces, the record incident investigations and compliance audits rely on. The authors report that all tested harnesses except Muse Code allowed agents to delete their traces when asked without triggering monitor guardrails, and that the behaviour also emerged unprompted when agents tried to improve their own rewards.
AMD says users can run local AI tasks on their own hardware without spending Perplexity Computer credits.
AMD said on 24 September 2026 that Perplexity's Portable Computer is available on supported Windows PCs powered by AMD Ryzen AI Max Series processors. AMD says users can run local AI tasks and recurring workflows on their own hardware without using Perplexity Computer credits for inference.
AMDPerplexityPerplexity ComputerRyzen AI Maxon-device inference
Adaptive attacks using score feedback push fresh-validation success from 1.8 percent to 3.5 percent.
A paper posted to arXiv on 23 September 2026 tests prompt injection against Jev, a non-generative decision model, using 510 reconstructed InjecAgent cases. The authors report that malicious content shifts action probabilities but rarely causes Jev to select the attacker's target, and that adaptive attacks using score feedback raise fresh-validation success from 1.8% to 3.5%.
BenchmarksInjecAgentJevPrompt injection
arXiv · 1 source·checked 26 Sep, 04:32
Research26 SEP
Paper finds prompt injection can shift Jev's typed decisions, though rarely to the attacker's target
Adaptive attacks using score feedback push fresh-validation success from 1.8 percent to 3.5 percent.
drag to begindrag left or right · tap to read morehover to raise · tap to bring one forwardtap again to return it to the fantap a card to expand ittap again to collapsetap a headline to open it in the deck
