正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Moondream 3 Preview adds frontier-level visual reasoning while retaining Moondream 2's blazing speed. No compromises.

moondream

11,866 subscribers

44,461 次观看 • 11 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

ARC-AGI-3 is out now! We've designed the benchmark to evaluate agentic intelligence via interactive reasoning environments. Beating ARC-AGI-3 will be achieved when an AI system matches or exceeds human-level action efficiency on all environments, upon seeing them for the first time. We've done extensive human testing that shows 100% of these environments are solvable by humans, upon first contact, with no prior training and no instructions. Meanwhile, all frontier AI reasoning models do under 1% at this time.

ARC-AGI-3 is out now! We've designed the benchmark to evaluate agentic intelligence via interactive reasoning environments. Beating ARC-AGI-3 will be achieved when an AI system matches or exceeds human-level action efficiency on all environments, upon seeing them for the first time. We've done extensive human testing that shows 100% of these environments are solvable by humans, upon first contact, with no prior training and no instructions. Meanwhile, all frontier AI reasoning models do under 1% at this time.

François Chollet

625,196 次观看 • 5 个月前

Step 2: Use 100+ models like GPT-5.1, Nano Banana Pro, Gemini 3, Claude 4.5... ✅ Expert-level reasoning, mathematics, and scientific knowledge ✅ Continuously updated with the latest models ✅ One account, no geographic restrictions Try GPT-5.1 for FREE now 👉

Step 2: Use 100+ models like GPT-5.1, Nano Banana Pro, Gemini 3, Claude 4.5... ✅ Expert-level reasoning, mathematics, and scientific knowledge ✅ Continuously updated with the latest models ✅ One account, no geographic restrictions Try GPT-5.1 for FREE now 👉

Leo ☄️

86,056 次观看 • 9 个月前

more on solana:2oKuzxekSwyGnqoiyJv5zVvBbRmh5UyRbjQ74j3Npump new launchpad in research preview on Robinhood (meow in replies if you want invite code) the mechanism, why I think it will change memes forever: 1. launch every new House starts with one hour to reach 0.05 ETH 2. level up level 1 → 0.1 ETH / 2 hours level 2 → 0.2 ETH / 4 hours ... level 8 → 12.8 ETH / 256 hours clear lv 8 & your House graduates into Uniswap pool 3. death miss the bar by the deadline? your House ends & liquidity dies into the King one level above else it dies into solana:2oKuzxekSwyGnqoiyJv5zVvBbRmh5UyRbjQ74j3Npump why this matters: - concentrates liquidity; no ghost LP - failed Houses feed stronger Kings or solana:2oKuzxekSwyGnqoiyJv5zVvBbRmh5UyRbjQ74j3Npump - every level is a PvP campaign to become King - fewer stronger memes goal is no wasted moni in dead pool, and one big fat cat contracts in research preview only need invites? meow in replies (follow retard.market to get DM)

more on solana:2oKuzxekSwyGnqoiyJv5zVvBbRmh5UyRbjQ74j3Npump new launchpad in research preview on Robinhood (meow in replies if you want invite code) the mechanism, why I think it will change memes forever: 1. launch every new House starts with one hour to reach 0.05 ETH 2. level up level 1 → 0.1 ETH / 2 hours level 2 → 0.2 ETH / 4 hours ... level 8 → 12.8 ETH / 256 hours clear lv 8 & your House graduates into Uniswap pool 3. death miss the bar by the deadline? your House ends & liquidity dies into the King one level above else it dies into solana:2oKuzxekSwyGnqoiyJv5zVvBbRmh5UyRbjQ74j3Npump why this matters: - concentrates liquidity; no ghost LP - failed Houses feed stronger Kings or solana:2oKuzxekSwyGnqoiyJv5zVvBbRmh5UyRbjQ74j3Npump - every level is a PvP campaign to become King - fewer stronger memes goal is no wasted moni in dead pool, and one big fat cat contracts in research preview only need invites? meow in replies (follow retard.market to get DM)

nicolette 🫪

66,180 次观看 • 1 个月前

Just dropped on HF — NeuTTS Air Next-gen on-device TTS that matches cloud-level quality while staying fully open source. > Real-time speech synthesis on CPU/GPU > 3-second voice cloning, no cloud or data upload > Compact: under 200 MB, runs on mobile and edge devices > Multilingual and expressive > Developed by Neuphonic , optimized for speed and fidelity

Just dropped on HF — NeuTTS Air Next-gen on-device TTS that matches cloud-level quality while staying fully open source. > Real-time speech synthesis on CPU/GPU > 3-second voice cloning, no cloud or data upload > Compact: under 200 MB, runs on mobile and edge devices > Multilingual and expressive > Developed by Neuphonic , optimized for speed and fidelity

steven

73,642 次观看 • 11 个月前

Moonshot AI is casually giving developers free daily access to Kimi K3 😳 no subscription no upfront payment just sign in and start using one of the largest open AI models available what you get for $0: - Kimi K3 with 2.8T parameters - 1M token context window - strong coding and reasoning performance - native vision capabilities - free daily credits that refresh automatically why this is worth checking: > access a frontier model without paying API fees > long context for large codebases and documents > works on web, desktop, mobile, and CLI getting started takes less than 2 minutes: 1. go to 2. create a free account 3. Kimi K3 is available as the default model 4. start chatting or coding with your daily free credits bonus: Moonshot Together lets you invite friends for a chance to earn 3, 7, 15, 30, or even 365 days of Kimi Membership through its rewards program benchmark highlights: > 2.8T parameter MoE model > 1M context window > strong performance across coding, browsing, and reasoning benchmarks important: free credits reset daily, rate limits apply on the free tier, and the open-weight release is expected on July 27 A simple way to try one of the latest frontier AI models without paying for API access

Moonshot AI is casually giving developers free daily access to Kimi K3 😳 no subscription no upfront payment just sign in and start using one of the largest open AI models available what you get for $0: - Kimi K3 with 2.8T parameters - 1M token context window - strong coding and reasoning performance - native vision capabilities - free daily credits that refresh automatically why this is worth checking: > access a frontier model without paying API fees > long context for large codebases and documents > works on web, desktop, mobile, and CLI getting started takes less than 2 minutes: 1. go to 2. create a free account 3. Kimi K3 is available as the default model 4. start chatting or coding with your daily free credits bonus: Moonshot Together lets you invite friends for a chance to earn 3, 7, 15, 30, or even 365 days of Kimi Membership through its rewards program benchmark highlights: > 2.8T parameter MoE model > 1M context window > strong performance across coding, browsing, and reasoning benchmarks important: free credits reset daily, rate limits apply on the free tier, and the open-weight release is expected on July 27 A simple way to try one of the latest frontier AI models without paying for API access

K2S

23,110 次观看 • 1 个月前

People wanna know why I love the SHREDmill so much?? This collection of videos is exactly the reason why I’m such a big advocate, as both a speed AND coaching/technical AMPLIFER⚡️🔌 All of these reps were done within the same session: Rep 1: 8.3 mph; excessive turnover, minimal horizontal push/thigh split Rep 2: after working on some positional wall drills, hit 7.7 mph; better positions/thigh split, but no push Rep 3: 9.3 mph; literally put it all together as far as the “good” from reps 1 and 2! Sprinting on the ground, incorporating max effort jumps/plyos, and still getting strong in the weight room are musts for maximizing speed development, but no other tool in the speed world allows you to both improve the qualities needed to be a damn good accelerator, while also giving you the feedback (both objectively via outputs, and the visual/auditory feedback from watching+coaching reps) needed to make improvements in short periods of time. Gettin’ fast FAST! Tony Villani SimpliFaster

People wanna know why I love the SHREDmill so much?? This collection of videos is exactly the reason why I’m such a big advocate, as both a speed AND coaching/technical AMPLIFER⚡️🔌 All of these reps were done within the same session: Rep 1: 8.3 mph; excessive turnover, minimal horizontal push/thigh split Rep 2: after working on some positional wall drills, hit 7.7 mph; better positions/thigh split, but no push Rep 3: 9.3 mph; literally put it all together as far as the “good” from reps 1 and 2! Sprinting on the ground, incorporating max effort jumps/plyos, and still getting strong in the weight room are musts for maximizing speed development, but no other tool in the speed world allows you to both improve the qualities needed to be a damn good accelerator, while also giving you the feedback (both objectively via outputs, and the visual/auditory feedback from watching+coaching reps) needed to make improvements in short periods of time. Gettin’ fast FAST! Tony Villani SimpliFaster

Cam Galgano, CSCS

11,766 次观看 • 4 个月前

I watched every 3-2 take by Bryce Harper in 2025. Here’s what it revealed: 1️⃣ Yes, Yes, No Approach - he is swinging until the ball tells him “no”. His barrel starts to turn on almost 100% of his takes. You will see these type of “active takes” with pretty much every high level hitter 2️⃣ He Can Stop Late - because his barrel turns rearward and creates depth, it allows him shut down or abort his swing as late as possible. This gives him a longer time to identify pitch speed, shape, and location

I watched every 3-2 take by Bryce Harper in 2025. Here’s what it revealed: 1️⃣ Yes, Yes, No Approach - he is swinging until the ball tells him “no”. His barrel starts to turn on almost 100% of his takes. You will see these type of “active takes” with pretty much every high level hitter 2️⃣ He Can Stop Late - because his barrel turns rearward and creates depth, it allows him shut down or abort his swing as late as possible. This gives him a longer time to identify pitch speed, shape, and location

Matt Antonelli

345,215 次观看 • 7 个月前

June 29 1960: A US Navy RA-3 crashes off Florida (US) all 3 crew die. While practicing landings on the carrier USS Saratoga, the arresting cable broke. With no means to stop safely, the plane went off deck. The crew attempted a go-around but were unable to, jet hit the sea at high speed and sank. The RA-3 was the photo-reconnaissance version of the A-3 “Skywarrior” bomber. The jet lost in the accident [144845] was attached to squadron VAP-62 (based at Fort. Lauderdale, Florida). The Jet´s wreckage was later discovered by chance by a film-making crew, who were producing a documentary on the Bermuda Triangle.

June 29 1960: A US Navy RA-3 crashes off Florida (US) all 3 crew die. While practicing landings on the carrier USS Saratoga, the arresting cable broke. With no means to stop safely, the plane went off deck. The crew attempted a go-around but were unable to, jet hit the sea at high speed and sank. The RA-3 was the photo-reconnaissance version of the A-3 “Skywarrior” bomber. The jet lost in the accident [144845] was attached to squadron VAP-62 (based at Fort. Lauderdale, Florida). The Jet´s wreckage was later discovered by chance by a film-making crew, who were producing a documentary on the Bermuda Triangle.

Francisco Cunha - ✈️ Aviation et al

385,832 次观看 • 2 个月前

50% more context unlocked for Qwen 3.8 27b Q4_K_XL dflash 2 on a single RTX 4090 (24 GB VRAM) I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B is now running a massive 250,000 context at 75 tokens/s on a single RTX 4090. Here is the secret: By default, `llama-server` reserves massive chunks of your VRAM to handle multiple concurrent users (batching). If you are running a single user session, you are bleeding memory for features you aren't using. By passing the `--parallel 1` flag, you force the engine to dedicate 100% of your 24GB VRAM buffer to a single user. When we combine the VRAM saved by our Q2_K 2-bit drafter with the VRAM saved by `--parallel 1`, the context ceilings absolutely explode: Note: all benchmarks carried out with a massive 28k prompt. Ubuntu 22. ### THE NEW 24GB PHYSICAL LIMITS (Single RTX 4090): # 1. The "Repo Swallower" (Q4 KV Cache): - Context: 250,000 tokens (Up from 170k!) - Speed: 73.66 t/s decode | 1,608 t/s prefill - Peak VRAM: 23.8 GB # 2. The "High-Precision SWE" (Q8 KV Cache): - Context: 150,000 tokens (Up from 100k!) - Speed: 75.01 t/s decode | 1,667 t/s prefill - Peak VRAM: 23.9 GB # 3. The "Pristine Attention" (Unquantized FP16 KV): - Context: 90,000 tokens - Speed: 80.58 t/s decode | 1,699 t/s prefill - Peak VRAM: 23.92 GB ### HOW TO RUN THE 250K GOD STACK TODAY: (Requires PR #27342 + my Q2_K Hugging Face drafter) llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q2_K.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 250000 -ngl 99 --parallel 1 --port 8080 -ctv q4_0 -ctk q4_0 We are pushing a quarter million tokens of context with speculative DFlash 2 decoding at 73 tokens/second on a single consumer gaming GPU. I dropped my custom 2 bit Hugging Face GGUF links, visual performance graphs, and the PR #27342 build instructions in the replies below. If you own a single RTX 3090 or 4090, it is officially time to cancel your API subscriptions and let local silicon eat the cloud. how much monthly API spend does an optimized 4090 rig like this actually replace for you?

50% more context unlocked for Qwen 3.8 27b Q4_K_XL dflash 2 on a single RTX 4090 (24 GB VRAM) I found a hidden VRAM tax in llama.cpp. By combining my custom 2 bit DFlash 2 drafter with one overlooked server flag, I just unlocked another +80,000 tokens of context. Qwen3.8-27B is now running a massive 250,000 context at 75 tokens/s on a single RTX 4090. Here is the secret: By default, `llama-server` reserves massive chunks of your VRAM to handle multiple concurrent users (batching). If you are running a single user session, you are bleeding memory for features you aren't using. By passing the `--parallel 1` flag, you force the engine to dedicate 100% of your 24GB VRAM buffer to a single user. When we combine the VRAM saved by our Q2_K 2-bit drafter with the VRAM saved by `--parallel 1`, the context ceilings absolutely explode: Note: all benchmarks carried out with a massive 28k prompt. Ubuntu 22. ### THE NEW 24GB PHYSICAL LIMITS (Single RTX 4090): # 1. The "Repo Swallower" (Q4 KV Cache): - Context: 250,000 tokens (Up from 170k!) - Speed: 73.66 t/s decode | 1,608 t/s prefill - Peak VRAM: 23.8 GB # 2. The "High-Precision SWE" (Q8 KV Cache): - Context: 150,000 tokens (Up from 100k!) - Speed: 75.01 t/s decode | 1,667 t/s prefill - Peak VRAM: 23.9 GB # 3. The "Pristine Attention" (Unquantized FP16 KV): - Context: 90,000 tokens - Speed: 80.58 t/s decode | 1,699 t/s prefill - Peak VRAM: 23.92 GB ### HOW TO RUN THE 250K GOD STACK TODAY: (Requires PR #27342 + my Q2_K Hugging Face drafter) llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -md Qwen3.8-27B-DFlash2-Q2_K.gguf --spec-type draft-dflash --spec-draft-n-max 3 -c 250000 -ngl 99 --parallel 1 --port 8080 -ctv q4_0 -ctk q4_0 We are pushing a quarter million tokens of context with speculative DFlash 2 decoding at 73 tokens/second on a single consumer gaming GPU. I dropped my custom 2 bit Hugging Face GGUF links, visual performance graphs, and the PR #27342 build instructions in the replies below. If you own a single RTX 3090 or 4090, it is officially time to cancel your API subscriptions and let local silicon eat the cloud. how much monthly API spend does an optimized 4090 rig like this actually replace for you?

Alok

39,189 次观看 • 15 天前

Qwen3.8-Flash-Next is starting to feel like the local model Opus fans have been waiting for. Someone ran the NVFP4 176B-class Flash-Next on 2× DGX Sparks, and the results are wild. Real measured scaling → C1: 44.2 tok/s → C2: 64.6 tok/s → C4: 86.8 tok/s aggregate The per-stream speed drops with concurrency, but total throughput keeps climbing. Long-context behavior was even more impressive: → 5K: needle retrieved → 21K: needle retrieved → 84K: needle retrieved → 167K: needle retrieved → 262K: prefill succeeded, but the window was saturated That 167K retrieval test is the one I care about. Long agent runs are where models usually start losing the plot. Flash-Next didn’t. It also held up surprisingly well on physics-heavy reasoning, artifact generation, research workflows, evidence checking, and long-horizon planning. The personality is interesting too. DeepSeek V4 Flash feels like the dependable workhorse. GLM-5.2 feels like the problem-solving machine. Qwen3.8-Flash-Next feels more insightful. It has that rare ability to understand what you’re actually asking rather than just following the surface pattern. The main weakness I’ve noticed is instruction following. It can occasionally drift between prose turns where DeepSeek and GLM stay tighter. And this is why the 256GB M5 Ultra conversation gets interesting. If Apple can pair that huge unified-memory pool with enough bandwidth, this model class becomes genuinely practical for long-running local agents. We’re talking about frontier-class reasoning on hardware sitting on a desk.

Qwen3.8-Flash-Next is starting to feel like the local model Opus fans have been waiting for. Someone ran the NVFP4 176B-class Flash-Next on 2× DGX Sparks, and the results are wild. Real measured scaling → C1: 44.2 tok/s → C2: 64.6 tok/s → C4: 86.8 tok/s aggregate The per-stream speed drops with concurrency, but total throughput keeps climbing. Long-context behavior was even more impressive: → 5K: needle retrieved → 21K: needle retrieved → 84K: needle retrieved → 167K: needle retrieved → 262K: prefill succeeded, but the window was saturated That 167K retrieval test is the one I care about. Long agent runs are where models usually start losing the plot. Flash-Next didn’t. It also held up surprisingly well on physics-heavy reasoning, artifact generation, research workflows, evidence checking, and long-horizon planning. The personality is interesting too. DeepSeek V4 Flash feels like the dependable workhorse. GLM-5.2 feels like the problem-solving machine. Qwen3.8-Flash-Next feels more insightful. It has that rare ability to understand what you’re actually asking rather than just following the surface pattern. The main weakness I’ve noticed is instruction following. It can occasionally drift between prose turns where DeepSeek and GLM stay tighter. And this is why the 256GB M5 Ultra conversation gets interesting. If Apple can pair that huge unified-memory pool with enough bandwidth, this model class becomes genuinely practical for long-running local agents. We’re talking about frontier-class reasoning on hardware sitting on a desk.

FHILY👑

19,992 次观看 • 9 天前

Just jumped into this 2016 Mercedes-Benz GLE 63 S AMG... and bro, this car is the most overprotective babe I've ever driven 😂 It's like it has trust issues with Nigerian roads! A lot happened while driving this car. 1. Another car edges too close? The steering assist just smoothly slides you away to a safe gap. No honking needed. It works like personal bodyguard. 2. If you are driving too close to a pavement, it drifts you away from the pavement on its own. 3. It senses real emergency (like sudden stop ahead)? Seatbelts tighten hard, front seats adjust, PRE-SAFE mode kicks in, and it brakes itself before you even react. First time e happen, my heart nearly left my body. But after a while... I started loving the extra care. Scary at first, but now? This level of pampering is sweet die. Luxury wey dey save life while you flex. Soup wey sweet na money kill am

Just jumped into this 2016 Mercedes-Benz GLE 63 S AMG... and bro, this car is the most overprotective babe I've ever driven 😂 It's like it has trust issues with Nigerian roads! A lot happened while driving this car. 1. Another car edges too close? The steering assist just smoothly slides you away to a safe gap. No honking needed. It works like personal bodyguard. 2. If you are driving too close to a pavement, it drifts you away from the pavement on its own. 3. It senses real emergency (like sudden stop ahead)? Seatbelts tighten hard, front seats adjust, PRE-SAFE mode kicks in, and it brakes itself before you even react. First time e happen, my heart nearly left my body. But after a while... I started loving the extra care. Scary at first, but now? This level of pampering is sweet die. Luxury wey dey save life while you flex. Soup wey sweet na money kill am

Verterex Properties 🏡 🚗

82,979 次观看 • 7 个月前

.Penn State Baseball ‘26 RHP Ryan Fuller (Avon Grove, PA) No. 7 RHP in PA | No. 120 RHP in USA 🇺🇸 at 6-2, 185 LBS. Fastball T95 at 2500 RPM with sufficient ride and run from 3/4 slot. Spins frisbee sweeper at 2600+ RPM while featuring sinker and high spin CH with 20in run. High and athletic leg drive into violent, whippy arm speed downhill. 𝗙𝗕: 𝟵𝟰-𝟵𝟱 (𝟮𝟱𝟬𝟬🔄 𝟭𝟳.𝟱𝗜𝗩𝗕, 𝟭𝟲𝗔𝗦𝗥 ↗️) 𝗦𝗞: 𝟵𝟯-𝟵𝟱 (𝟮𝟯𝟴𝟱🔄 𝟴.𝟱𝗜𝗩𝗕, 𝟭𝟵𝗔𝗦𝗥 ➡️) 𝗖𝗛: 𝟴𝟳-𝟴𝟴 (𝟮𝟰𝟯𝟬🔄 𝟯𝗜𝗩𝗕, 𝟮𝟬𝗔𝗦𝗥 ➡️) 𝗦𝗪𝗣: 𝟴𝟯 (𝟮𝟲𝟬𝟱🔄 -𝟱𝗜𝗩𝗕, -𝟭𝟲𝗛𝗠 ↙️) #MLKProDay26 | Ascent Athlete Prep Baseball Pennsylvania | Ryan Fuller

.Penn State Baseball ‘26 RHP Ryan Fuller (Avon Grove, PA) No. 7 RHP in PA | No. 120 RHP in USA 🇺🇸 at 6-2, 185 LBS. Fastball T95 at 2500 RPM with sufficient ride and run from 3/4 slot. Spins frisbee sweeper at 2600+ RPM while featuring sinker and high spin CH with 20in run. High and athletic leg drive into violent, whippy arm speed downhill. 𝗙𝗕: 𝟵𝟰-𝟵𝟱 (𝟮𝟱𝟬𝟬🔄 𝟭𝟳.𝟱𝗜𝗩𝗕, 𝟭𝟲𝗔𝗦𝗥 ↗️) 𝗦𝗞: 𝟵𝟯-𝟵𝟱 (𝟮𝟯𝟴𝟱🔄 𝟴.𝟱𝗜𝗩𝗕, 𝟭𝟵𝗔𝗦𝗥 ➡️) 𝗖𝗛: 𝟴𝟳-𝟴𝟴 (𝟮𝟰𝟯𝟬🔄 𝟯𝗜𝗩𝗕, 𝟮𝟬𝗔𝗦𝗥 ➡️) 𝗦𝗪𝗣: 𝟴𝟯 (𝟮𝟲𝟬𝟱🔄 -𝟱𝗜𝗩𝗕, -𝟭𝟲𝗛𝗠 ↙️) #MLKProDay26 | Ascent Athlete Prep Baseball Pennsylvania | Ryan Fuller

Chris Dryll

19,968 次观看 • 7 个月前

There is not one snap from Saturday that perfectly encapsulates Ohio States defense and what Arch was dealing with. Pre-snap: 4-2-5 MOFC look, which means nothing since Patricia drops into Tampa/C2 from this look often. Post-snap: Double A gap Creeper (Mahomes) with MOFC. Versus 3x1 Gun Weak creates 4th/1st and 3 Up rules. - Boundary 5t is the 4th/1st Crosser - Field 5t is the Strong Hook - Caleb Downs is the 3 up is 3 player - Nickel is SCIF/B2b player. Caleb Downs gets #3 vertical to start the down that turns into an inside stem. The post safety that starts into the boundary identifies #3 bending back over the middle of field, and Caleb Downs is able to "Cut" him back to the post safety. Essentially the 3 up is 3 and post safety switch roles in them middle of the play. This adds layers to the coverage based off of leverage (#3 is running away from O/s leverage 3 up player, while running into postive leverage post safety). All of this is to say this new Ohio State defense is not predicated on just being talented and being chaotic, they showed an extreme grasp on techniques just in week 1.

There is not one snap from Saturday that perfectly encapsulates Ohio States defense and what Arch was dealing with. Pre-snap: 4-2-5 MOFC look, which means nothing since Patricia drops into Tampa/C2 from this look often. Post-snap: Double A gap Creeper (Mahomes) with MOFC. Versus 3x1 Gun Weak creates 4th/1st and 3 Up rules. - Boundary 5t is the 4th/1st Crosser - Field 5t is the Strong Hook - Caleb Downs is the 3 up is 3 player - Nickel is SCIF/B2b player. Caleb Downs gets #3 vertical to start the down that turns into an inside stem. The post safety that starts into the boundary identifies #3 bending back over the middle of field, and Caleb Downs is able to "Cut" him back to the post safety. Essentially the 3 up is 3 and post safety switch roles in them middle of the play. This adds layers to the coverage based off of leverage (#3 is running away from O/s leverage 3 up player, while running into postive leverage post safety). All of this is to say this new Ohio State defense is not predicated on just being talented and being chaotic, they showed an extreme grasp on techniques just in week 1.

The Ballmanac

20,153 次观看 • 1 年前

Seedance 2.0 Sensitive Content Guide – Part 2 🔥 This morning we covered the fundamentals. Tonight we put them into practice and successfully generated Giant Ant-Man vs Spider-Man. Here’s exactly how we did it: • Used extremely detailed physical descriptions only (no character names) • Ant-Man described as “a giant hero in advanced red & black high-tech combat suit with insect-shaped helmet” • Spider-Man described as “an agile hero in red & blue suit with complex black web patterns” • Action described cinematically (“power and speed collision”, “intense energy impact”) • Clean 3-shot JSON structure with controlled pacing This method is currently one of our most reliable ways to combine multiple major characters in one scene. Full working prompt is attached below. Want more advanced examples and stable prompts? Comment “PART 3” and we’ll drop the next level. We’re building this system together.

Seedance 2.0 Sensitive Content Guide – Part 2 🔥 This morning we covered the fundamentals. Tonight we put them into practice and successfully generated Giant Ant-Man vs Spider-Man. Here’s exactly how we did it: • Used extremely detailed physical descriptions only (no character names) • Ant-Man described as “a giant hero in advanced red & black high-tech combat suit with insect-shaped helmet” • Spider-Man described as “an agile hero in red & blue suit with complex black web patterns” • Action described cinematically (“power and speed collision”, “intense energy impact”) • Clean 3-shot JSON structure with controlled pacing This method is currently one of our most reliable ways to combine multiple major characters in one scene. Full working prompt is attached below. Want more advanced examples and stable prompts? Comment “PART 3” and we’ll drop the next level. We’re building this system together.

Clonizer

30,434 次观看 • 6 个月前

Today we're launching a new experiment. Over the next 30 days, Dave will become one of the internet's most recognizable AI creators. Dave is just a potato. That's the point. We want to demonstrate that something as simple as a potato can become a global media personality with the right AI workflows, creative direction, and distribution strategy. We'll launch one platform at a time, gradually increasing the level of AI automation and sophistication while comparing the results. Day 1: X - simple reactions, almost no AI. Day 2: Instagram. Day 3: TikTok. Day 4: YouTube. Day 5: Threads. Everything is built almost entirely with open-source models in ComfyUI. Follow Dave:

Today we're launching a new experiment. Over the next 30 days, Dave will become one of the internet's most recognizable AI creators. Dave is just a potato. That's the point. We want to demonstrate that something as simple as a potato can become a global media personality with the right AI workflows, creative direction, and distribution strategy. We'll launch one platform at a time, gradually increasing the level of AI automation and sophistication while comparing the results. Day 1: X - simple reactions, almost no AI. Day 2: Instagram. Day 3: TikTok. Day 4: YouTube. Day 5: Threads. Everything is built almost entirely with open-source models in ComfyUI. Follow Dave:

Kevin Raheja

36,481 次观看 • 1 个月前

This guy made $400K with his own script on Polymarket Yes, it’s true, the dude wrote a script using ClawdBot and now it earns him passive income No programming knowledge No insider tools for building the script Just a solid script that delivers results Profile → Copytrade → When I opened the code, I was genuinely surprised No massive databases No insanely complex infrastructure Nothing rocket-science level His FULL strategy: 1. “Low-Risk NO Plays” The bot targets near-impossible outcomes and stacks small, high-probability wins. This isn’t gambling, it’s closer to systematic risk underwriting. 2. Logical Price Gaps If event A must logically lead to event B and the market hasn’t priced that in yet, the bot enters instantly. While you’re still reading the headline, the imbalance is already gone. No human can compete with that speed. 3. Main Advantage - Sports and Politics These markets are flooded with retail money and delayed or emotional reactions. The bot lives in the spread, clipping tiny profits from every mismatch. Scale Effect Tens of thousands of micro-trades every month, each worth just a few cents. Over time they compound into seven-figure results. Final Take There’s a real bot war happening on Polymarket. Crypto markets are already slowed down by fees and heavy competition, but sports markets are still pure chaos, easy money for those who automate.

winkle.

57,003 次观看 • 7 个月前

Qwen 3.8 27B (dense) running on a single RTX 4090 (24GB VRAM) at 65 tokens/sec decode with MTP! 260,000 context window or 65 tokens/sec decode with native MTP. The API cartel should be terrified. We are officially running frontier tier agentic AI (benchmarks comparable to claude opus 4.6 max) on a single consumer gaming GPU. I benchmarked Qwen3.8-27B on a single NVIDIA RTX 4090 (24GB VRAM, Ubuntu 22) using Unsloth’s Dynamic Q4_K_XL GGUF on the latest llama.cpp. Here is the complete benchmark breakdown across both Context Scaling and MTP Overdrive (28k prompt baseline): ### PART 1: The Context Scaling Matrix (Pure Throughput) # 1. Standard FP16 KV Cache (Unquantized): - 80k Context: 2,664.7 t/s prefill | 40.68 t/s decode | 22.36 GB VRAM - 100k Context: 2,678.7 t/s prefill | 40.89 t/s decode | 23.59 GB VRAM (100k is the hard ceiling for unquantized f16 KV in 24GB VRAM) # 2. Q8 Quantized KV Cache (-ctv q8_0 -ctk q8_0): - 130k Context: 2,639.1 t/s prefill | 40.96 t/s decode | 22.18 GB VRAM - 170k Context: 2,653.9 t/s prefill | 40.70 t/s decode | 23.68 GB VRAM (170k is the sweet spot for heavy agentic coding workflows) # 3. Q4 Quantized KV Cache (-ctv q4_0 -ctk q4_0): - 260k Context: 2,659.8 t/s prefill | 40.70 t/s decode | 23.00 GB VRAM Full 262k native context residing entirely in 24GB VRAM. Zero system RAM offload. Stress test with a monster 142k real-world prompt (-c 170000, Q8 KV): - Prefill: 1,829.50 tokens/s - Decode: 31.3 tokens/s - VRAM: 23.7 GB rock solid ### PART 2: Native MTP Overdrive (Trading Context for Speed) Since MTP heads are baked into the architecture, enabling native speculative drafting pushes decode speeds straight to 60 t/s with zero external draft model: # 1. MTP + Q8 KV Cache: - 80k Context: 2,370.66 t/s prefill | 59.25 t/s decode | 23.4 GB VRAM (MTP state buffers eat slightly more memory, making 80k the ceiling for Q8) # 2. MTP + Q4 KV Cache: - 130k Context: 2,391.09 t/s prefill | 60.10 t/s decode | 23.5 GB VRAM (Sweet spot: 130,000 context running at a screaming 60 tps decode) ### Qwen3.8-27B vs Muse Glimmer 30B Two days ago I benched Meta's Muse Glimmer 30B hitting 130k context unquantized (19.3 GB VRAM) pulling 50-75 t/s decode. If you own a single RTX 3090 or RTX 4090, you have zero excuse to burn API credits. ### The Reproduction llama.cpp flags: 1. Max Context Stack (260,000 Context @ 41 tps): ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -c 260000 -ngl 99 --port 8080 -ctv q4_0 -ctk q4_0 2. MTP Overdrive Stack (130,000 Context @ 60 tps): ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -c 130000 -ngl 99 --port 8080 -ctv q4_0 -ctk q4_0 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Unsloth's Hugging Face GGUF links, intelligence/agentic benchmark details, and performance charts are posted in the replies. Local compute is eating the cloud alive. How much monthly API spend does a 24GB setup like this actually replace for you?

Qwen 3.8 27B (dense) running on a single RTX 4090 (24GB VRAM) at 65 tokens/sec decode with MTP! 260,000 context window or 65 tokens/sec decode with native MTP. The API cartel should be terrified. We are officially running frontier tier agentic AI (benchmarks comparable to claude opus 4.6 max) on a single consumer gaming GPU. I benchmarked Qwen3.8-27B on a single NVIDIA RTX 4090 (24GB VRAM, Ubuntu 22) using Unsloth’s Dynamic Q4_K_XL GGUF on the latest llama.cpp. Here is the complete benchmark breakdown across both Context Scaling and MTP Overdrive (28k prompt baseline): ### PART 1: The Context Scaling Matrix (Pure Throughput) # 1. Standard FP16 KV Cache (Unquantized): - 80k Context: 2,664.7 t/s prefill | 40.68 t/s decode | 22.36 GB VRAM - 100k Context: 2,678.7 t/s prefill | 40.89 t/s decode | 23.59 GB VRAM (100k is the hard ceiling for unquantized f16 KV in 24GB VRAM) # 2. Q8 Quantized KV Cache (-ctv q8_0 -ctk q8_0): - 130k Context: 2,639.1 t/s prefill | 40.96 t/s decode | 22.18 GB VRAM - 170k Context: 2,653.9 t/s prefill | 40.70 t/s decode | 23.68 GB VRAM (170k is the sweet spot for heavy agentic coding workflows) # 3. Q4 Quantized KV Cache (-ctv q4_0 -ctk q4_0): - 260k Context: 2,659.8 t/s prefill | 40.70 t/s decode | 23.00 GB VRAM Full 262k native context residing entirely in 24GB VRAM. Zero system RAM offload. Stress test with a monster 142k real-world prompt (-c 170000, Q8 KV): - Prefill: 1,829.50 tokens/s - Decode: 31.3 tokens/s - VRAM: 23.7 GB rock solid ### PART 2: Native MTP Overdrive (Trading Context for Speed) Since MTP heads are baked into the architecture, enabling native speculative drafting pushes decode speeds straight to 60 t/s with zero external draft model: # 1. MTP + Q8 KV Cache: - 80k Context: 2,370.66 t/s prefill | 59.25 t/s decode | 23.4 GB VRAM (MTP state buffers eat slightly more memory, making 80k the ceiling for Q8) # 2. MTP + Q4 KV Cache: - 130k Context: 2,391.09 t/s prefill | 60.10 t/s decode | 23.5 GB VRAM (Sweet spot: 130,000 context running at a screaming 60 tps decode) ### Qwen3.8-27B vs Muse Glimmer 30B Two days ago I benched Meta's Muse Glimmer 30B hitting 130k context unquantized (19.3 GB VRAM) pulling 50-75 t/s decode. If you own a single RTX 3090 or RTX 4090, you have zero excuse to burn API credits. ### The Reproduction llama.cpp flags: 1. Max Context Stack (260,000 Context @ 41 tps): ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -c 260000 -ngl 99 --port 8080 -ctv q4_0 -ctk q4_0 2. MTP Overdrive Stack (130,000 Context @ 60 tps): ./build/bin/llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -c 130000 -ngl 99 --port 8080 -ctv q4_0 -ctk q4_0 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-p-min 0.7 Unsloth's Hugging Face GGUF links, intelligence/agentic benchmark details, and performance charts are posted in the replies. Local compute is eating the cloud alive. How much monthly API spend does a 24GB setup like this actually replace for you?

Alok

377,759 次观看 • 22 天前

This is one of the wildest AI products I've seen this year. GeoAxis takes any photo and returns its exact location in seconds with no GPS and metadata needed. How it works: Most images online have the metadata stripped. Screenshots, reposts, cropped photos, none of them carry GPS coordinates. So the location has to be recovered from the pixels alone. This model reads several visual signals in parallel: → Landmarks (buildings, statues, signage) → Vegetation (species, density, seasonality) → Architecture (window style, roof pitch, materials) → Terrain (elevation cues, coastline, mountain profile) → Sky and lighting conditions Each signal narrows the search space. Vegetation alone can pin you to a climate zone. Architecture can pin you to a region. A single readable sign or a distinctive skyline can pin you to a city block. The pipeline runs coarse-to-fine: 1. Geo-Estimation: rough country/region prediction 2. Precise Geolocation: meter-level match against their indexed image database 3. Indoor Geolocation and Car Identification are shipping next Claimed accuracy is meter-level in deployed cities, with billions of images indexed. For OSINT, journalism and reverse image search, this changes the workflow. For personal privacy, stripping metadata no longer does anything.

Hasan Toor

15,030 次观看 • 15 天前

If you're an emerging manager or solo GP in 2026, you have four moats mega-funds literally cannot replicate. This is the best time in 20 years to be small. 1. Concentration math. A $5B fund needs $1B exits to move the needle, while a $50M fund returns the fund on a $100M exit. The top-decile 2024 PE returns (71.70%) came from concentrated bets, while diversified portfolios got the median. Mega-funds literally can't run 15-name portfolios because the math doesn't clear. 2. Founder access at seed. A $5B fund can't write $500K seed checks. The economics don't work and the LP base wouldn't tolerate it. Emerging managers can. The funds that own seed positions in 2026's eventual top-decile names will outperform every mega-fund that came in at Series B. Access compounds. You can't manufacture it later. 3. Decision speed. Mega-funds have IC structures, LP reporting cycles, and legal review processes that turn 48-hour decisions into 6-week ones. Emerging managers can wire in two days. In a market where the best deals close in a week, two-day speed is the difference between getting in and reading about it. 4. Public transparency. Most mega-funds have legal and LP constraints that limit what they can say publicly. Emerging managers don't. They can publish their thesis, share their portfolio reasoning, post their fund construction math. That transparency is the cheapest fundraising channel ever invented, and the funds using it now are compounding credibility before they need it. Three of these are yours by default. The fourth is what fundraises your next vintage.

If you're an emerging manager or solo GP in 2026, you have four moats mega-funds literally cannot replicate. This is the best time in 20 years to be small. 1. Concentration math. A $5B fund needs $1B exits to move the needle, while a $50M fund returns the fund on a $100M exit. The top-decile 2024 PE returns (71.70%) came from concentrated bets, while diversified portfolios got the median. Mega-funds literally can't run 15-name portfolios because the math doesn't clear. 2. Founder access at seed. A $5B fund can't write $500K seed checks. The economics don't work and the LP base wouldn't tolerate it. Emerging managers can. The funds that own seed positions in 2026's eventual top-decile names will outperform every mega-fund that came in at Series B. Access compounds. You can't manufacture it later. 3. Decision speed. Mega-funds have IC structures, LP reporting cycles, and legal review processes that turn 48-hour decisions into 6-week ones. Emerging managers can wire in two days. In a market where the best deals close in a week, two-day speed is the difference between getting in and reading about it. 4. Public transparency. Most mega-funds have legal and LP constraints that limit what they can say publicly. Emerging managers don't. They can publish their thesis, share their portfolio reasoning, post their fund construction math. That transparency is the cheapest fundraising channel ever invented, and the funds using it now are compounding credibility before they need it. Three of these are yours by default. The fourth is what fundraises your next vintage.

fabrizio

25,335 次观看 • 4 个月前