Загрузка видео...

Не удалось загрузить видео

На главную

200HP vs 1000HP Speed Comparison 🌆🚗🌃🏎️ Longer version on YT which shows the 'Stiff' Camera mode(less FOV, less movement) #screenshotsaturday #gamedev #indiedev #gamedesign #jdm #indiegame

58,210 просмотров • 2 лет назад •via X (Twitter)

Комментарии: 10

Фото профиля Night-Runners
Night-Runners2 лет назад

Main song: ASC - Binary

Фото профиля Savachi
Savachi2 лет назад

you did a really great job with that camera and fov especially. this comprasion really feels like it has difference. i remember all past race games released, and their 300km/h doesn't feel at all. but yours 280~... it gave me goosebumps just like in childhood

Фото профиля 1991 Nissan Figaro
1991 Nissan Figaro2 лет назад

what i think would be kinda cool is if we can put our mp3 files in the game and use them as a custom soundtrack. BUUT im no game dev, and copyright may be involved 😅 game looks good tho, dont rush yourself pls <3

Фото профиля Silver
Silver2 лет назад

This game looks so sick

Фото профиля M̷. ɐɯᴉɹnl
M̷. ɐɯᴉɹnl2 лет назад

reminds me of Tokyo Xtreme Racer 3

Фото профиля i n f i n i d r i f t 💨
i n f i n i d r i f t 💨2 лет назад

the sense of speed is amazing😍

Фото профиля stef826
stef8262 лет назад

I've seen a fair bit of gameplay and this game seems really cool tbh, but I'm wondering if it will ever come to console

Фото профиля rico suave
rico suave2 лет назад

@plaguedme

Фото профиля phil
phil2 лет назад

when can i play this im BEGGING

Фото профиля ba
ba2 лет назад

Can't wait for the releases

Похожие видео

First, the conclusion: the so called Galaxy Z Fold8 Ultra leak from last night turned out to be a false alarm. After watching the original video over and over again, I'm convinced it isn't a Galaxy Z Fold8 Ultra. It's actually the blue version of the Galaxy Z Fold7. Because of the lighting and color temperature, the blue frame appeared black on camera, which led to the confusion. Here are the three biggest clues: 1. The inner display camera cutout is too large. Compared with the official Galaxy Z Fold8 Ultra renders, the inner camera hole on the Fold8 Ultra is noticeably smaller. The one in the leaked video is still the larger Fold7-style cutout. 2. The rear camera housing doesn't match. The cylindrical camera housing reflects light with a bright metallic sheen, which is much more consistent with the blue Galaxy Z Fold7 than the black Galaxy Z Fold8 Ultra. Some of the hardware details also don't line up. 3. The inner display is still highly reflective. The device in the video shows obvious screen reflections. However, the Galaxy Z Fold8 Ultra is the first Fold to feature an anti reflective inner display film. Anyone familiar with Samsung foldables should immediately notice that the reflections don't look like this. Based on these three points, I'm confident the device in the video is a blue Galaxy Z Fold7, not the Galaxy Z Fold8 Ultra. The confusion is understandable because the two generations are incredibly similar. They share the same overall dimensions and aspect ratio, making them very easy to mix up. The good news is that the real Galaxy Z Fold8 Ultra will be even better. It will have an even less visible crease than the Fold7, along with the new anti reflective inner display film.

Ice Universe

126,235 просмотров • 21 дней назад

DeepSeek allegedly beat ChatGPT. So should we all switch? Here's a comparison for academics: 👇 The prompt: 🎯 I asked a fairly complicated question on how to analyze my data collection to gain novel ecological insights and for references to similar papers for each suggestion. Here are the results: 1️⃣ Content: - DeepSeek provided deeper, shorter replies that I perceived as more creative. - ChatGPT played it safe and gave me slightly more obvious solutions 2️⃣ Hallucinations: - DeepSeek found 40+ search results and each cited paper had a link attached. No hallucinations. - ChatGPT hallucinated 2 out of 7 papers. 3️⃣ Speed: - DeepSeek (free version) was many times slower and often didn't reply due to server overload 4️⃣ Customizability: - ChatGPT's replies are much more customisable. By using "Custom Instructions" you can vary the length and include more or less formulae. 5️⃣ Length: - ChatGPT generally gave much longer and more detailed replies. 5️⃣ Guidance: - ChatGPT created a 4-step plan on using my data, which made sense but was quite vague on implementation topics. - DeepSeek gave me links to the coding packages I should be using, which was very helpful. 🤔 My 2 cents If you're a premium subscriber to ChatGPT, keep using it - the results are very similar and can be fine-tuned. Keep an eye out DeepSeek, as it is much more efficient (~10x cheaper under the hood), it might produce more impressive models later on.

Ilya Shabanov

38,452 просмотров • 1 год назад

🎬 Whimsical Reverie | Preview: Dawning Heartlight >> Main Attribute: Elegant 🦁 Outfit Details [Dawning Heartlight] can be evolved into the following outfits: [Dawning Heartlight: Empty Cocoon] (Stage 1), [Dawning Heartlight: Falling Flame] (Stage 2), and [Dawning Heartlight: Renewed Gleam] (Stage 3). Fully glow it up to obtain the [Dawning Heartlight: Fledgling] outfit. 🦁 Outfit Ability After using the ability, Nikki summons a golden-maned lion, which enters Follow Mode. Nikki can freely switch between four modes: Follow, Mount, Hover, and Sprint. While in Mount Mode, she can switch to Sprint Mode, greatly increasing movement speed. Use the Hover button to leap into the air and enter Hover Mode. While hovering, use the ability button to select a target and dive toward it. Press the Hover button again to end Hover Mode and land. In Ability Settings, Smart Mount Mode can be enabled. When enabled, pressing the ability button will switch directly between mounting and dismounting without entering Follow Mode. Mount via Interact can also be enabled, allowing the Interact button to be used for mounting and dismounting. Even after switching to another ability, the Interact button can still be used to mount or dismount. When disabled, mounting can only be controlled with the ability button. ❖ Special Effect: In the open world, after Nikki fires a charged arrow while riding the golden-maned lion, the lion will let out a roar, knocking back Esselings within a certain area ahead and briefly slowing Whimstars within a certain surrounding area. Note: The [Gilded Dawnchaser: Mount] ability cannot be used in Realm Challenges or any boss stages. Hover Mode is unavailable in uncleared dungeons. 🦁 Resonance Rewards Based on your number of Resonance attempts in [At Dreams' Dawn], you can earn rewards such as [Dawning Heartlight: Makeup], [Momo's Cloak: Radiant Wish], and the decoration [Sacred Heartyard]. This decoration can be placed in open spaces for photography and display. 🦁 Evolution Rewards Reaching Evolution Stage 1 unlocks the pose [Wings Renewed]. Reaching Evolution Stage 2 unlocks [Feather's Wish] for Momo's Camera: Framed Moment and the decoration [Sacred Heartyard: Radiance], which features a new look, special effect colors, and interactive animations. Reaching Evolution Stage 3 unlocks [Wishful Wind Ballad: Makeup], a new idle animation, the poses [Longing for the Light], [Weaving Dreamlight], and [Echoes Within], plus enhanced special effects and alternative details for certain pieces. —— ➤ Infinity Nikki Version 2.8 Golden Dust launches globally on July 16 (UTC-7)! ➤ Download Now: ➤ Join us on Facebook Group: #InfinityNikki #GoldenDust #GoldenCity

Infinity Nikki

58,554 просмотров • 22 дней назад

♾Infinity Nikki Version 2.8 Outfit Preview: White Night's Ballet and Dawning Heartlight Klick, klick... She hears the mechanical heart beating— Golden feathers whirl, a flash of golden mane breaks through the mirror, and there, she sees the soul of a god aflame— In the fractured frame, a miracle arrives at last: the moment she meets herself once more. ♾ [White Night's Ballet] >> Main Attribute: Cool >> Outfit Ability: [Spinning Reflections] After using the [Spinning Reflections] ability, Nikki bursts through the shattered mirror and begins an elegant dance amid blooming shards of light, with her movement speed greatly increased. If approaching a wall, Nikki can slide over the wall's surface and continue moving seamlessly. Sliding continuously consumes energy. When energy is depleted, Nikki will exit the sliding state. Wall sliding consumes more energy. While sliding across a wall, Nikki can jump over obstacles. She can also enter the floating state to disengage from the wall. When using the ability during [Companion Journey], Nikki and her companion will have a special movement effect. After equipping the [Spinning Reflections] ability, Nikki can also summon the Lucent Mirror and Mirror Shadow and activate the Brightnight Mirror Realm. The Brightnight Mirror Realm can be placed on water, allowing Nikki to slide within it. Within the Brightnight Mirror Realm, select [Airy Dance] from the sidebar to switch movement posture. Approach and remain near a Mirror Shadow to light it up. Interact with the lit Mirror Shadow to change its outfit, check its styling, try on its outfit, or invite friends to try it on. ❖ The ability [Spinning Reflections] cannot be used in uncleared dungeons. The Brightnight Mirror Realm cannot be placed in certain stages or in the Mira Crown Merit Arena. >> [Ocean's Radiance Gift Box] The box includes the item [Dance of Golden Days] and the dance move [Flourishing Steps]. The item [Dance of Golden Days] can be used to dance gracefully in the open world. The dance move [Flourishing Steps] can be used in Momo's Camera. ❖ Available after reaching 160 Resonances in the limited-time Resonance [Through Songbound Night]. >> Evolution Rewards ◆ Evolving the outfit to Stage 1 unlocks the pose [Graceful Leap]. ◆ Evolving the outfit to Stage 2 unlocks the Framed Moment item [Eternal Ballet]. You can also use Finale Heartchime in [Clothing Evolution > Action: Illusory Dance] to obtain the item [Dance of Golden Days: Illusory Dance] and the dance move [Flourishing Steps: Illusory Dance]. ◆ Evolving the outfit to Stage 3 unlocks [The Final Dance: Makeup], a new idle animation, poses, enhanced special effects, and alternative details for certain pieces. ♾ [Dawning Heartlight] >> Main Attribute: Elegant >> Outfit Ability: [Gilded Dawnchaser: Mount] After using the ability, Nikki summons a golden-maned lion, which enters Follow Mode. Nikki can freely switch between four modes: Follow, Mount, Hover, and Sprint. While in Mount Mode, she can switch to Sprint Mode, greatly increasing movement speed. Use the Hover button to leap into the air and enter Hover Mode. While hovering, use the ability button to select a target and dive toward it. Use the Hover button again to end Hover Mode and land. In Ability Setup, Smart Mount Mode can be enabled. When enabled, using the ability button will switch directly between mounting and dismounting without entering Follow Mode. Mount via Interact can also be enabled, allowing the Interact button to be used for mounting and dismounting. Even after switching to another ability, the Interact button can still be used to mount or dismount. When disabled, mounting can only be controlled with the ability button. ❖ Special Effect: In the open world, after Nikki fires a charged arrow while riding the golden-maned lion, the lion will let out a roar, knocking back Esselings within a certain area ahead and briefly slowing Whimstars within a certain surrounding area. ❖ The [Gilded Dawnchaser: Mount] ability cannot be used in Realm Challenges or any boss stages. Hover Mode is unavailable in uncleared dungeons. >> Decoration: [Sacred Heartyard] Can be placed in open spaces. Follow the lingering song of the wind in your memories and climb the steps of a dream. Soul Spring murmurs softly, guiding the way to where flowers bloom in abundance. ❖ Available after reaching 160 Resonances in the limited-time Resonance [At Dreams' Dawn]. >> Evolution Rewards ◆ Evolving the outfit to Stage 1 unlocks the pose [Wings Renewed]. ◆ Evolving the outfit to Stage 2 unlocks the Framed Moment item [Feather's Wish]. You can also use a Radiance item in [Clothing Evolution > Decoration: Radiance] to obtain the decoration [Sacred Heartyard: Radiance]. ◆ Evolving the outfit to Stage 3 unlocks [Wishful Wind Ballad: Makeup], a new idle animation, poses, enhanced special effects, and alternative details for certain pieces. ♾ Follow Infinity Nikki and retweet this post, and we'll select 10 lucky stylists to win [Infinity Nikki Aeroplane Chess Set *1] each. —— ➤ Infinity Nikki Version 2.8 Golden Dust launches globally on July 16 (UTC-7)! ➤ Download Now: ➤ Join us on Facebook Group: #InfinityNikki #GoldenDust #GoldenCity

Infinity Nikki

77,065 просмотров • 22 дней назад

I just compared Claude Code vs Codex vs Cursor CLI The task was to build a Next.js app with Tailwind 4 and shadcn components to collect customer feedback and showcase it with a widget. I gave all three the same prompt and let them go for 30 minutes to see what they came up with. Claude Code with Opus 4.1 Even though I told it to set up the app in the existing project folder, it tried to create a directory for it. After I interrupted and told it not to do that, it built a demo form and landing page with no errors. I had to ask it to make the demo interactive so users could submit a testimonial and preview it. The landing page looked like AI and was pretty basic, but it worked and it was done in a fraction of the time of the others. Total tokens used: 33k Codex with GPT-5 At the end of the 30 minutes I just could not get Codex to produce a working app. It got stuck in a loop of not being able to set up Tailwind 4 and despite many, MANY, attempts, I ended up with a "failed to compile" error. Total tokens used: 102k Cursor Agent with GPT-5 This was the slowest agent by far and a couple of times I actually thought it got stuck in a loop and was close to Ctrl+C'ing to cancel it. The TUI is really nice though, especially how it shows diffs and it did eventually build a working app (after one or two slight errors that needed fixing) The demo was interactive and it had a very minimal design that looked bare but also a lot less like an "AI generated" app than the Opus 4.1 design. It also wasn't too chatty and just did what it needed to do! Code quality was on a par with Opus 4.1, but it did use 5.5x as many tokens to get there. Still cheaper than Opus on a direct comparison but not when you factor in a Claude Code Max subscription. Total tokens: 188k I'll be able to do a proper comparison and record some videos when I'm back from holiday but for now, Opus is still the more capable model out of the box and Claude Code is the more complete CLI product. It will be interesting to see how Cursor evolve their CLI though with commands and subagents because I think with GPT-5 they have a real shot at providing competition for Claude Code if they can optimise output to get similar quality with less tokens. Jump to 0:40 in the video to see the two apps. Which do you think is which? ;)

Ian Nuttall

194,949 просмотров • 1 год назад

Assassin's Creed Black Flag Resynced | New Gameplay Details: ▫️ Parkour has been improved for faster, more fluid movement ▫️ Edward now gets a short speed boost when mantling after a wall run ▫️ You can also roll upon hitting the ground to maintain a sprinting pace ▫️ Ziplines have been added ▫️ Advanced Parkour options are back, enabling a deeper level of control, including Manual Jumping ▫️ Observe has been added from AC Shadows - basically a stronger version of Eagle Vision that tags enemies, loot, and quest objectives ▫️ Edward can now crouch at any time during stealth making him less visible ▫️ Enemies take longer to spot you in the dark ▫️ Use an arsenal of tools - Blowdarts (Sleep or Berserk), Smoke Bombs, Rope Dart, and more ▫️ You can raise or lower your hood at any time ▫️ Tailing and eavesdropping missions no longer result in a descynchronization when falling out of range ▫️ The Hiring Dancers return, allowing Edward to recruit a group of dancers as a distraction on the streets ▫️ Edward can blend into any group of three or more civilians ▫️ Throw money to create chaos in the streets and lure guards to a location ▫️ The core combat loop revolves around triggering takedowns ▫️ Hidden Blade takedowns, environmental takedowns, and perfect parry finishers ▫️ Environmental combat - kick enemies off ledges, blow up powder kegs, or drop enemies to the ground for a takedown ▫️ Edward now has a Perfect Dodge attack ▫️ Edward now has a charged Heavy Attack ▫️ The Gun Kata returns, allowing you to take down multiple enemies in quick succession ▫️ Releasing July 9th on PS5, Xbox, and PC #AssassinsCreed #AssassinsCreedBlackFlag #BlackFlagResynced

KAMI

70,295 просмотров • 2 месяцев назад

Watched "The Black Book" on Netflix last night. The Nollywood film is called Nigeria's version of "John Wick". It's been viewed >70m times and topping charts eveywhere (peaked at #3 for English globally and top 10 in 69 countries). The best part: it cost only $1m to make. While that budget is significant for Nollywood films — which are typically made for less than $50,000 — it's a drop in the bucket for Hollywood blockbusters. Here is a random ratio list for the "first 28-day watches" vs. budget for major Netflix original films: ◻️ "The Gray Man" (139m watch / $200m budget = 0.7x) ◻️ "Red Notice" (230m / $200m budget = 1.2x) ◻️ "Don't Look Up" (171m / $75m = 2.3x) ◻️ "Glass Onion" (136m / $40m = 3.4x) ◻️ "Bird Box" (156m / $20m = 7.8x) "The Black Book" is huge bang for the buck at 70x watches-to-budget. The film was written and directed by former tech and marketing exec Editi Effiong. To get funding for the film — and make Hollywood-calibre action scenes — Effiong tapped entrepreneurs from Nigeria's tech ecosystem. Legendary Nollywood actor Richard Mofe-Damijo plays the revenge-seeking protagonist. Nollywood is currently the world's second largest film industry by volume, making >2,000 a year (this is ~2x the US market and trails only India's Bollywood). Piracy is rampant in Nollywood, so big budgets for in-country distribution doesn’t make sense (rather, quality over quantity wins the day). But the success of "The Black Book" on streaming can clear the path for more global Nollywood hits My suggestion for another Keanu-themed Nollywood banger: “Speed”.

Trung Phan

650,668 просмотров • 2 лет назад

Some time ago, I had the idea to port NVIDIA Physical AI stack to AMD. The motivation was to improve hardware diversity and enable world models and VLAs to run beyond a single ecosystem. We started with NVIDIA Cosmos Predict 2.5-2B. Porting wasn’t trivial: these models are deeply optimized for NVIDIA’s stack. We used this as an opportunity to apply our ROCm kernels. The results were surprising: Both encode and diffusion run faster on AMD Instinct MI300X vs. NVIDIA H200 (FA3) and we still saw significant headroom for further optimization. Quality is unchanged across modalities (validated with WorldJen) To be clear, this is no luck. We have deep experience with diffusion models and AMD GPUs. But this just gives us a good opportunity to get closer to a true hardware-to-hardware comparison, as we work with less software abstractions than usual. Just to give an example, on AMD, memory instructions are async with a hardware queue of ordered pending instructions, enabling concurrent load/store with compute without warp specialization. Bottom line: there are real architectural advantages on AMD, if you take the time to work with the hardware. Note, we did tradeoff ~20% higher memory usage, That being said, AMD has more to give to begin with :) in the coming weeks: AMD versions of Cosmos Transfer and GR00T, an even faster version of Cosmos Predict, and open-sourcing an attention kernel faster than AITER v3 (which is closed-source for some reason? cc: Anush Elangovan )

Omer Shlomovits

36,620 просмотров • 3 месяцев назад

The #GTA 3 port to the Sega #Dreamcast has been progressing at an incredible pace. It's been amazing to see the whole DC community come together to tag-team this "impossible" project... Here it is running on a stock DC, no longer requiring the 32MB RAM hardware modification, only a few weeks into development. Since I finally got the time to sit down, build the codebase, and look into some of what I think is the critical path for performance, let's talk about some technical shit, and some of the future steps I think can be taken to further improve performance. First of all, I want everyone to take note that this is NOT a port of the PS2 version. This is a port of the PC version, which has extra content, increased draw distance, improved textures, and other things that have actually increased the challenge here... Whether the DC version will ultimately have these additions or not will remain to be seen, but we're running into plenty of shit that the PS2 didn't have to worry about (like these big-ass PC replay saves won't fit onto a Visual Memory Unit!) Secondly, lets talk about what is and isn't currently optimized, because it's absolutely vital that the DC's hardware is fully utilized here for the sake of performance and achieving a competitive polygon count. Unlike with modern devices, where the whole graphics pipeline is handled by the GPU, both the PS2 and Dreamcast were responsible for transforming and doing lighting calculations for each vertex BEFORE they got submitted to the GPU. The PS2 had a vector coprocessor to do this, while the Dreamcast had a few extremely important SIMD and fast math assembly instructions on its CPU to do these computations. Up until literally just a few hours ago (not shown in this footage), the Dreamcast's SH4 was doing 100% of these operations in slow-ass plain C and C++ code, which is absolutely sub-optimal and is immediately bogging down its CPU with just transforming vertices, bottlenecking the entire graphics pipeline on the fist T&L stage, and also leaving less CPU time for handling other gameplay logic... this is going to absolutely have to be addressed (and already has begun to be). Another issue that is crippling performance here is the fact that the models are all using individual triangles rather than triangle strips, which the Dreamcast's PVR GPU was designed to handle better... Converting these models to use strips rather than individual triangles will result in MANY different gainz for the DC, as you're going from 3N to N+2 vertices per triangle. Converting the models to triangle strips will 1) reduce load times, since model assets will be smaller 2) reduce the amount of video memory required to hold these vertices on the GPU 3) reduce the amount of shit that must be transferred from the CPU to the GPU and 4) give us back a bunch of CPU time, since the SH4 will be less bogged down transforming redundant vertices! TL;DR: This is still EXTREMELY suboptimal in terms of fully utilizing the graphical potential of the Dreamcast. There is going to be a LOT that can be done still to both improve performance and polygon counts, so stay tuned! FINALLY: Mad respect and love to Stefanos Kornilios Mitsis Poiitidis, for doing an amazing job leading this project, and to Frogbull , Esppiral, and everyone else who is helping us stick it to the PS2 by making this happen! #gamedev #retrogaming #cplusplus

Falco Girgis

88,356 просмотров • 2 лет назад

This is the best scene in Hell Grind, an entirely AI-made movie, the flashback. Watch it and read this analysis on where we are with AI movies today: time, cost, quality. Overall: Phenomenal technical demo by Higgsfield. Mediocre movie. Good graphics, hints of emotion, but superhero movie level quality in certain scenes at best. Too many cuts. That said, 660x fewer man hours, 50x faster and 36x cheaper than the median US film. Time: The 95 min film took 15 people 14 days. The median US theatrical production takes ~200 people ~2yrs. That’s a 660x improvement in man-hours and 50x in calendar time. Economics: It took $500k, 80% of which was compute. The final footage was cut from ~100hrs of footage generated from text to video / image to video models like Bytedance’s Seedance: a 64:1 “curation” ratio. The median US movie takes ~$18M, with even indie films costing $1-5M. Thats 36x cheaper than median. Quality: Average watch *at best*. Way too many cuts between shots, several characters change accents and have “AI” synthetic voices and characters feel like it’s AI too. Movement, editing and blocking feel artificial too. On the plus side, we’ve more or less solved character consistency, camera angles and realism. The reason the movie wasn’t amazing was more about poor directorial choices than innately unusable video models. Hard to put a number on it but maybe we’re at ~90% on quality that is technically achievable. If Scorsese made an AI movie, I reckon it would be quite good. I know the visceral reaction to anything AI is real and well-studied. But I think it’s folly to fight the inevitability of AI film. It’s too cheap and quick to ignore and almost there on quality. Creators with distribution *will* make AI films and shows and just put them on YouTube. This is the worst quality, slowest and most expensive it will ever be. In the end, good content beats “real” content.

Deedy

35,098 просмотров • 1 месяц назад

Scaling campaigns overseas sounds like a creative problem. Honestly, it’s not. The real bottleneck is localization. As a product lead, I’ve lost too many weeks waiting for native voice actors, rebuilding region-specific edits, and manually fixing lip-sync issues that still looked slightly off in the final export. The worst part is that every new market turns into another production branch to maintain. That simply does not scale. So over the last few weeks, I started testing a few different AI localization workflows with our own ecommerce video ads to see which ones could actually survive real production conditions. Wizstar_official ended up being the one we kept coming back to. Not because it generated the flashiest demo. Because it was the first one that consistently held together once we pushed it into actual multi-market production. The Video Translation workflow supports 12 major languages, which already covers most global consumer markets we care about. But what stood out during testing was how natural the localization sounded. Not “translated”. Actually localized. The tone, pacing, and delivery felt native enough that most people on our team genuinely stopped noticing it was AI-generated after a few runs. More importantly, the video itself stays intact. Audio and visual timing remain aligned after translation, lip-sync holds even during side angles and faster speech, and multi-character scenes stay surprisingly stable instead of collapsing into mismatched cuts. That matters a lot more in production than benchmark style demos. We also tested it against a few other tools internally, and Wizstar consistently handled complex scenes better, especially when multiple speakers, product close-ups, and fast pacing were involved. The output needed significantly less cleanup before going live. Video Reference was another reason we kept using it. Being able to reuse existing high-performing ecommerce structures instead of rebuilding creative logic market by market saves an unreasonable amount of time. Seedance 2.0 also supports face input and multi-model orchestration, which noticeably improves character consistency and scene stability across longer sequences. After a few projects, Wizstar quietly became part of our workflow. We can now produce localized ecommerce creatives in minutes instead of rebuilding entire pipelines around every market. If you want to test it yourself: New users get free credits on signup. First subscription is $19 and includes a complimentary 30-second Ecommerce Agent workflow to test features like Product to Video. Let the tools handle the production overhead. The side-by-side comparison below shows one of our English masters translated into Spanish while keeping almost the exact same pacing and vibe intact. #Wizstar #AIVideo #GrowthHacking

Leo Reed

121,626 просмотров • 2 месяцев назад

I’ve been using GPT-5.6 Sol internally for the past two months, I've spent probably 25+ billion tokens. Here’s my review and comparison to Fable 5: > Let's start with the analogy because everyone seems to be giving theirs - GPT-5.6 is likely the last version of the GPT-5 training run series. It's kind of like an athlete at their peak. Through years of experience in the game, they've become the most reliable player and has the highest game IQ. But, there's no more room to grow. Fable on the other hand, being essentially the first version of a new training run, is the first round draft pick rookie. Raw talent mixed with the energy only a young person would have results in some incredible plays we didn't think possible, but also mistakes due to lack of experience. But that rookie will only improve and likely will be better than the veteran ever was because it's a new game and a new era. > GPT-5.6 is genuinely better at long, sustained work. With /goal, I've had it running complex projects for days with almost no intervention. It built a Minecraft-style game, kept adding features and mobs after the core game worked, and only stopped because I stopped the run. I never felt as though I had to jump in and guide it back to the right path. > It keeps finding useful work when you give it a concrete finish line. I had it recreate Excel with a loop. It inspected the real desktop excel app with Computer Use, comparing that against its own build, and closing the gaps. I stopped it after six days after it had built an incredible amount of functionality. > It's faster than other models in two different ways. The raw generation speed is higher, something OpenAI has been putting effort into. But it also takes a shorter path to solutions. It wanders less, changes less code, and generally knows how to get things done directly. In daily use, it feels about 2-3x times faster than Fable. That's my impression, not a controlled benchmark. The difference is large enough that I notice it constantly. > It works well across a wide range of tasks. I use it for one-line edits, quick questions, browser chores, and multi-day builds without changing my prompting style. Speaking of browser control, its the best ever I've used. To the point where I actually use it often. If a task lives on a website, GPT-5.6 usually opens the browser and does it there instead of asking for an API key or forcing everything through the terminal. When I switched back to GPT-5.5, it went straight to the command line even when the browser was clearly the better tool. > And it can handle real browser work, not just toy demos. During a data import, I had it monitor Supabase and resize instances as the load changed. It stayed on the dashboard, adjusted capacity, and checked the result without an API or a custom script. > I also gave it a full Google Workspace migration. It moved Forward Future from to preserved the old aliases, and configured MX, SPF, and DKIM. Before a consequential save, it stopped, explained exactly what would change, and waited for confirmation. > The reasoning setting matters a lot. Light is good for questions and small edits. High and Extra High are the sweet spots for serious work. Ultra usually takes longer than the extra thinking is worth and burns tokens. > I love that 5.6 is split into 3 sizes. Not only can you control speed and cost that way, but you still also have the thinking effort setting for each of them. Very precise controls. I just wish Codex automatically routed my prompts for me. > Its personality is blunt and a little bland. Claude feels warmer and more natural to talk to. GPT-5.6 is more clinical, but I like that for work. It gives me enough explanation and rarely pads the answer. I usually have to ask Fable to explain things more simply and/or more concise. > Its front-end taste has improved, but the default is predictable. Left alone, it turns websites into PowerPoint decks with huge statements and hard section breaks. The good news is that it takes design direction well and can revise without destroying the parts that already work. > It still makes confident mistakes. I asked it to rebuild parts of a system, and it told me the job was finished. Later, I found out it wasn't. Bits of its internal process also leak into the answer occasionally. > Claude Fable is more naturally autonomous on large, open-ended projects. GPT-5.6 is easier to reach for. I don't need to invent a huge project to justify using it. It works just as well for a small edit or browser chore. > GPT-5.6 is also cheaper. Sol costs $5 per million input tokens and $30 per million output tokens. Fable costs $10 and $50. Cached input is cheaper too. Still, cost per finished task matters more than cost per token. > GPT-5.6 isn't the best at everything, and it still needs supervision. But it generates faster, wanders less, works at almost any scale, and wastes less of my time. It's the model I have the most confidence in to get the job done right the first time. I put together a full breakdown with all the tests, prompts, and examples on a site. You can read it here:

Matthew Berman

186,712 просмотров • 25 дней назад

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

Avi Chawla

269,667 просмотров • 1 месяц назад

The Oil Shock Is Hitting A Supply Chain Already Rewired Under Stress The real danger here is hat the global economy is entering the worst energy shock of our lifetimes after already spending the last year rewiring trade under tariff pressure, China decoupling, front loaded inventories, and longer supply routes. That matters because fragility rarely comes from one shock. It comes when one shock lands on top of a system that already had its buffers removed. The China Trade Collapse The China trade collapse is real. Census data shows U.S. goods imports from China fell from $438.7 billion in 2024 to $308.4 billion in 2025, a decline of almost 30%. U.S. exports to China fell from $143.2 billion to $106.3 billion, and the bilateral goods deficit dropped to $202.1 billion, the lowest level in years. That is not a blip. That is structural decoupling. But the key nuance is that global trade did not collapse with it. UNCTAD says global trade still grew about 7.5% in 2025 to a record $35 trillion, while BEA data shows total U.S. imports rose 4.8% and the overall trade deficit barely moved, falling only 0.2% to $901.5 billion. The Rerouting Illusion That means the world did not stop trading. It rerouted. China was partly replaced by Mexico, Vietnam, India, Taiwan, Thailand, Indonesia, and other nodes. On paper, that looks like resilience. In reality, it also means more complexity, more shipping dependence, more customs friction, more insurance risk, more working capital tied up in inventory, and more fuel burned per unit of goods moved. The supply chain survived the tariff shock by becoming less efficient. Now Place An Oil Shock On Top Of That The IEA says global observed oil inventories fell by 85 million barrels in March, while stocks outside the Middle East Gulf fell by 205 million barrels as Hormuz flows were choked off. It also said oil export losses now exceed 13 million barrels per day, with cumulative supply losses of more than 360 million barrels in March and 440 million projected for April. The peak daily supply loss is already above 12 million barrels per day, larger than the 1973 Arab oil embargo at roughly 4.5 million barrels per day and the 1978 to 1979 Iranian Revolution at roughly 5.6 million. The Inventory Signal The chart is showing the physical cushion disappearing. Once inventories fall fast enough, price stops being the only rationing mechanism. The system starts rationing through behavior, policy, and scarcity. Airlines cut routes. Truckers pass through diesel costs. Food prices rise. Fertilizer gets tighter. Refineries prioritize. Governments release reserves. Then come pressure campaigns, fuel allocation, export controls, industrial curtailments, remote work mandates, speed limits, and priority access for military, emergency services, farming, and food logistics. What An Energy Lockdown Would Actually Look Like That is what an energy lockdown would look like. Not necessarily COVID style house arrest, but a forced reduction in mobility because the fuel system cannot support normal economic life at normal prices. The 1970s gave us gas lines, odd even rationing, Sunday station closures, and a 55 mph speed limit. Today’s version would be more technocratic, more targeted, and probably sold as temporary conservation. But the logic is the same. When energy is scarce, freedom of movement becomes a policy variable. My Take The oil shock is bad enough by itself. The real danger is that it is hitting a trade system already made more fragile by tariffs and China decoupling. Trade did not collapse in 2025. It rerouted. But rerouted trade depends on cheap, available fuel. If Hormuz stays disrupted, this becomes an inflation, logistics, food, credit, and political stability shock. The market is focused on the price of oil, but the real warning is the inventory draw. Once the spare barrels are gone, energy stops being managed by markets alone and starts being managed by allocation.

EndGame Macro

41,818 просмотров • 3 месяцев назад

Time for a dose of reality. This 👇is the supposedly smoking gun video that Rep. Anna Paulina Luna said last night had "never been seen before" and "could blow open the entire JFK investigation." Luna said she learned about it from Oliver Stone and said the video "shows Oswald near the vehicle [JFK's limo] when the assassination took place, which means he couldn't have been the shooter." And she told a credulous Jesse Watters, NBC had been "very, very much guarding this tape." A small problem. It is NOT a secret video that has never been seen. It is available for anyone to watch at the Sixth Floor Museum in Dallas (or on their YouTube channel). What Rep. Luna is referring to is a clip of a film taken by Jimmy Darnell, an NBC affiliate cameraman who had been riding in the "press camera cars" behind the motorcade. He, and three other cameramen had jumped out of their car after the shots and started filming about 30 seconds AFTER the assassination. Darnell's film shows Dallas police officer Marrion Baker running to the front door of the Texas School Book Depository, and past the TSBD's supervisor, Roy Truly, both of whom would encounter Oswald in less than another 30 seconds, on the second floor. Oswald was on his way out of the building, having come down from the sixth floor after the shooting. Truly vouched for Oswald and Baker let him go. Some conspiracy theorists think that a few seconds in the Darnell film shows a figure on the far-left side of the front of the Texas School Book Depository they believe is Oswald. The figure has been dubbed prayer man, some say because of the way he has his arms crossed in front of him, while others think the conjecture is the equivalent of a conspiracy Hail Mary prayer long shot. For those not into the weeds on the JFK assassination, this is not the first time that a film or photo has been used to try and exonerate Oswald. For many years, conspiracy theorists contended that Oswald was visible in one of the pictures taken by James Altgens, an AP photographer at the scene. Oswald, they contended, was the fuzzy figure in a white t-shirt standing to the left entrance of the TSBD. That was in fact one of Oswald's coworkers, Billy Lovelady. Even AFTER Lovelady identified himself as the person in the photo, some conspiracy theorists refused to believe him. Oliver Stone and others have now turned to the Darnell film and the person standing to the left of Lovelady. Maybe that is Oswald. It is not even clear from the fuzzy image if it is absolutely a white man. It could even be a light-colored black or Latin person. But to Stone and Luna and others, it is Oswald. Thirteen of Oswald’s work colleagues were standing at the front steps of the Texas School Book Depository. They were there to watch the motorcade. Not one of them remembers seeing Oswald there. That has not stopped the conspiracy theorists, who claim the clothing on the unidentified man looks like what Oswald wore when he was arrested after killing Dallas police officer J.D. Tippit. Ten other Texas School Book Depository employees wore shirts like the one on “prayer man.” The entire exercise might as well be a Rorschach inkblot test. That is what happened before in the case, for instance, with a half-inch square portion of Mary Moorman’s badly faded Polaroid taken a split second after the assassination. Conspiracy theorists enhanced it and blew it up and thought they had found the image of the phantom second shooter on the Grassy Knoll, someone they dubbed “badgeman.” They claimed that greatly enlarged pixels identified a rifle and a Dallas police uniform. The enhanced photo only shows shadows and trees, no person, no shooter. Now, Oliver Stone and others believe the only thing preventing an ironclad identification of Oswald in the Darnell film is that they do not have the original first-generation. The Sixth Floor Museum version is a second-generation print of the film that the Museum acquired as part of a personal collection in 2006. No one is certain if a first-generation version exists, but the guess by Luna and Stone is that it must be somewhere in the long-lost archives at NBC. Not quite the same as NBC having, as Rep. Luna claimed, “very, very much guarding it.” I hope that the Oversight Committee gets the original Darnell video. Many of us would like to see the clearest possible version of the frenetic scenes that played out in the minutes after the assassination. For those, however, who think it will be dispositive and exculpatory evidence that Oswald was hanging out at the front of the School Book Depository, good luck. That story might earn a headline in the National Enquirer or the Daily Mail. It should not be the ‘breaking news’ from the congresswoman leading the new investigation into the assassination of President Kennedy.

Gerald Posner

173,327 просмотров • 1 год назад