Is it Google's responsibility to ensure employees eat well in office cafes? Food director Michiel Bakker says "we're not the food police" but can be "architects" of healthy choices on menus at #WSJFutureFest# #FutureWork#

This new Nvidia paper is huge for model routing and the future of inference: re-using the KV cache across different LLMs. One of the greatest challenge in model routing during long-horizon agent settings is that the cache does not transfer across different models. As a result, naively switching models without KV-cache-awareness will result in wasting more money. Conversely, transferring the KV cache across models significantly expands the total amount of savings that can be achieved. This paper is a great step in that direction. Congrats to Taekyung Heo, Rasoul Shafipour, Ritchie Zhao, Maximilian Golub, @MMKamani7, @Ritika_Borkar, Makesh Tarun Chandran, Pantea Zardoshti, and Bita Darvish Rouhani. Two questions: 1. Do you plan to investigate cross-family transfer? I saw this in the future work section but curious to learn more about your intuitions on the opportunities and challenges here. 2. What was the reason for focusing mostly on small to large model transfers and not large to small?

We shouldn't be chained to the past—our past shouldn't dictate the quality of our future work. Instead, all our future innovations and right decisions should actually help resolve past issues as well. #notetomyself#

Claude usagemaxxing guide: - minimize use of subagents -> try to coordinate work in the main agent - don't use xhigh or max reasoning -> use high - don't compact sessions -> start new sessions for new tasks - don't break cache -> maintain a HANDOFF.md file that tracks future work for new sessions to kick off

SHOCKING EVIDENCE FOUND BY APPLE AGAINST OPENAI OF STEALING ITS DATA The company has accused OpenAI and two former Apple employees of stealing its confidential data as OpenAI ventures into consumer hardware. Both employees Chang Liu and Tang Yew Tan held senior positions at Apple. The case was filed on July 10, but the biggest evidence came only after they got access to Liu's MacBook issued by Apple on August 21. A forensic check allegedly found that Liu: 1. Accessed a confidential Apple circuit schematic and used it to run simulations at OpenAI. 2. Trained an AI agent to run the simulator and review results on its own. 3. Asked an OpenAI colleague to destroy evidence after learning about Apple’s investigation. 4. Used the same tool at OpenAI that Apple uses internally for its own work Apple was only able to gather this evidence because Liu had used the file on his Mac mini, which synced via iCloud to the same MacBook he'd kept, creating a forensic trail it could trace. Apple says the AI agent use is particularly concerning. Once confidential data enters an AI system, its learnings could be used in future work making the damage difficult to fully undo. Overall 400 former Apple employees now work at OpenAI which is why Apple is calling it a coordinated scheme operating “at every level” to obtain its proprietary data. Apple now wants the court to speed up evidence gathering, give access to Liu's Mac mini and stop OpenAI from using any of its trade secrets. OpenAI has denied everything and wants the case dismissed. The next hearing is scheduled for October 1. But if Apple proves its claims, this could become one of the biggest trade secret battles in tech.

DNA origami nanorobots selectively eliminate tumor cells while protecting normal tissue. Researchers from Sweden’s Karolinska Institutet engineered a DNA-based nanorobot capable of destroying cancer cells without substantially harming healthy surrounding tissue. This microscopic device transports a potent cytotoxic molecule that stays concealed until it reaches the acidic conditions typical of many solid tumors. Upon detecting the lower pH, the nanorobot uncovers its lethal payload and initiates targeted destruction of malignant cells. Laboratory tests confirmed that the weapon stays dormant at physiological pH yet activates strongly in the acidic milieu associated with tumors. In mouse models of breast cancer the active nanorobot curtailed tumor expansion by roughly 70 percent relative to an inert control version. Construction relies on DNA origami, a method that precisely folds DNA strands into nanoscale architectures programmed for defined functions. The integrated safety mechanism aims to minimize collateral injury to healthy cells, addressing a persistent limitation of conventional anticancer therapies. Still at an early developmental stage, the platform has not undergone human trials. Future work will examine performance in more complex cancer models and refine tumor-type specificity. Successful progression could enable a novel class of precision cancer treatments. ["Nanorobot with hidden weapon kills cancer cells." Karolinska Institutet]

I had an incredible opportunity last week. Great scope, right up my alley. For a large enterprise based here in Australia. Ultimately I had to turn it down. Why? Cost. The available budget wasn't enough for me to believe that I could deliver the quality that the customer would expect. I had two options. Undercook the outcome or take a hit and hope for future work. Neither seemed viable to me. I don't want to deliver poor outcomes and I want a sustainable studio so ultimately I had to say no. Right move?

TL;DR: Falcon, a 0.5M-parameter model, does streaming audio-visual target speaker extraction using lip video, with about 20x less compute than AV-TF-GridNet. The paper also releases RealSSA, a more realistic evaluation set. Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes Key points ・🎧 Existing benchmarks mostly use two speakers and anechoic mixes, missing reverberation and near/far-field interference ・🏠 RealSSA-Sim covers four room types with 1–5 near-field and 10–15 far-field interferers (about 153 hours), plus a real phone-recorded test set ・🧩 Falcon is a single-pass U-Net separator that models time-frequency only at the bottleneck and injects global context into skip connections via GCI ・🎯 A training-only speaker contrastive loss pushes outputs away from near-field interferers, with no added inference cost ・⚡ 10.6G MACs and 51.3 ms GPU latency, versus 45.3G and 172.1 ms for SwiftNet-12 ・📈 Best SI-SNRi and SDRi on all four simulated benchmarks, though AV-Mossformer2 edges it on real-recording WER It balances efficiency and accuracy while making evaluation more realistic. The loss needs clean interferer waveforms, so applying it to real recordings is future work. #SpeechProcessing# #AudioVisual#