Researchers traverse across Antarctica in minus 50 degree wind chill to validate measurements from NASA’s Ice, Cloud and Land Elevation Satellite. “Cold, dry air fills your nose and lungs. Your eyes blink, trying in vain to bat away the overwhelming brilliance of sunlight reflecting off the snow,” field notes described.

As Goldman bankers traversed the world finding investors to buy SpaceX stock at its blockbuster IPO, a fleet of the bank's wealthy clients were preparing to sell their early investments in Elon Musk’s company.

TDN-1 attack drone on its first (piloted) test flight from Traverse City, Michigan, United States, 19 May 1943. The pilot was Navy Lieutenant C.C. Corley. The TDN-1 was a World War II era American assault drone developed by the Naval Aircraft Factory for the US Navy. It was one of the earliest operational attack drones and a precursor to the better known TDR-1 built by Interstate Aircraft. The program grew out of the Navy's "Project Option," an effort led by Rear Admiral Delmar Fahrney to develop television guided unmanned aircraft that could deliver torpedoes or bombs against heavily defended targets without risking a pilot. The TDN-1 was designed in 1942 and first flew in 1943. It was a twin engine, mid wing monoplane built largely of non strategic materials such as wood, with fixed tricycle landing gear to simplify mass production and keep critical aluminum free for combat aircraft. It was powered by two Lycoming O-435-2 air cooled engines of about 220 horsepower each. The drone could carry a 2,000 pound bomb or a Mark 13 aerial torpedo slung beneath the fuselage. A small cockpit allowed a pilot to ferry the aircraft manually, but in combat configuration it was meant to fly unmanned. Control came from a "mother" aircraft, typically a TBM Avenger fitted with an RCA Block television receiver and a radio command link. A small TV camera in the drone's nose transmitted a live picture back to the controller, who steered it onto the target with a joystick. Maximum speed was roughly 140 miles per hour, with a range of about 425 miles. It was slow and not especially robust, which limited its tactical usefulness. Around 100 airframes were built, mostly at the Naval Aircraft Factory in Philadelphia and at a Navy run plant in Clearwater, Florida. The TDN-1 was used mainly for training drone controllers and developing tactics. Operational combat missions were carried out by the improved TDR-1 with Special Task Air Group One (STAG-1) in the Solomon Islands in 1944, where assault drones struck Japanese ships and shore positions before the program was cancelled later that year as conventional air power proved sufficient. Although it never saw front line combat itself, the TDN-1 helped prove that television guided, radio controlled strike aircraft were technically feasible, foreshadowing later cruise missiles and modern attack drones.

DeepSeek V4.1 Gives Prefill and Decode Different Compute Paths Prompt tokens mostly traverse 20 layers; generated tokens traverse all 40. This Causal Encoder-Decoder design nearly halves long-input prefill while preserving autoregressive generation. Zhihu contributor 潜龙勿用, Changxin Ke(柯昌鑫), a graduate researcher at ICT, CAS, explains how its architecture and post-training were designed together. 1️⃣ A causal encoder, not T5 The 40-layer backbone is split into a 20-layer causal encoder and a 20-layer decoder. Both remain causal. The encoder processes the prompt and supplies the decoder’s global KV. Most prompt tokens avoid the decoder stack, while generated tokens still run through all 40 layers. That means 8B active parameters per prefill token versus 16B per decode token. 2️⃣ Most layers share global memory CSA2 uses three modes: Full creates global KV and an index; Reindex shares the KV but selects new positions; Reuse shares both. Only four layers create independent global KV, four reindex it, and 30 reuse both KV and the latest index. Each layer retains its own query and local SWA state. The first Full layer builds up to 16,384 candidates. Later layers search this pool for their Top-512. From 4K to 1M context, decode FLOPs per token rise only about 25%. 3️⃣ Serving approximations enter training Exact reconstruction of decoder SWA states would replay 2,560 prompt tokens. V4.1 replays only the final 128 encoder outputs and trains the model to tolerate the approximation. With FP4 global KV, cache falls to 890 bytes per token, persistent cache to roughly one eighth of V4-Flash, and prefill compute close to half. Bounded replay and constrained retrieval are not last-minute serving tricks. The model experiences them during post-training. 4️⃣ Post-training is an evolving Agent system Each RL task combines a problem, environment, and verifier. New trajectories can reveal shortcuts, broken environments, or verifier errors and send the task back for repair. V4.1 trains across multiple harnesses, merges checkpoints between RL runs, and finishes with on-policy distillation from more than 40 teachers. Raising reasoning effort from 25 to 100 increases output length about 2.5×, while average Pass@1 across eight benchmarks rises from 67.1% to 76.3%. ✅ The real design choice V4.1 aligns model structure, cache policy, retrieval limits, training environments, and inference around long-running Agents. The model learns under the same constraints the deployed system will actually impose. 🔗 Full analysis: #DeepSeek# #DeepSeekV41# #LLMArchitecture# #AIAgents# #LongContext# #AIInfra#