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。

Check out the actual speed (not yet the final version) of Qwen3-Next-80B-A3B-Instruct on Apple MLX! 🔥 4-bit: 67 TPS 8-bit: 58 TPS bf16: 48 TPS Movie normal speed, only waiting times removed. Awni Hannun and Gökdeniz Gülmez did it and I bet there is still room for improvement 💪

Ivan Fioravanti ᯅ

21,560 subscribers

15,946 次观看 • 1 年前 •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

MTPLX V2.10 is out! I put a lot of work into this one. Highlights: - Qwen 3.8 Next Support Added: Bare Speed (4 bit) and Optimised Speed (4 bit/8 bit) models added running at 50 - 80 TPS sustained at high contexts. (expect speed improvements in the coming days) - Qwen 3.8 Next mmap: The 32 GB n-gram table streams from SSD via mmap instead of RAM. Both models fit a 96 GB Mac with 0 speed impact. - Qwen 3.8 27B Decode & Prefill boost EVERYWHERE: +15% TPS at 3k context, +29% at 88k, +54% at 147k. Prefill +41%. - 48 GB Macs went from 3 to 4 TPS in swap @ 30k context to 33 TPS. - KV Cache has been solved!: 8 bit and 4 bit cache now barely impacts speed instead of a 50% decline. - Coding agents wall time optimisations: file task dropped from 150s to 44s, and mid-session first token went from ~2s to 0.11s. - CLI UX improvements + 34 bug fixes. My thoughts on Qwen Next? Incredible model. Enormously better at design, and creative tasks with scary good vision. It created what I think is the best flappy bird. maybe even better than the Fable version. It had various difficulties and skins, moving pillars, and upgrades. It also had the most memeified look. Incredibly impressed. However, it is slightly worse than 27B at having code just work straight away. But it's personality is also far better than 27B. Here is the Flappy bird it made below! Made with Qwen Next Optimized Speed one shot@ 57 TPS:

MTPLX V2.10 is out! I put a lot of work into this one. Highlights: - Qwen 3.8 Next Support Added: Bare Speed (4 bit) and Optimised Speed (4 bit/8 bit) models added running at 50 - 80 TPS sustained at high contexts. (expect speed improvements in the coming days) - Qwen 3.8 Next mmap: The 32 GB n-gram table streams from SSD via mmap instead of RAM. Both models fit a 96 GB Mac with 0 speed impact. - Qwen 3.8 27B Decode & Prefill boost EVERYWHERE: +15% TPS at 3k context, +29% at 88k, +54% at 147k. Prefill +41%. - 48 GB Macs went from 3 to 4 TPS in swap @ 30k context to 33 TPS. - KV Cache has been solved!: 8 bit and 4 bit cache now barely impacts speed instead of a 50% decline. - Coding agents wall time optimisations: file task dropped from 150s to 44s, and mid-session first token went from ~2s to 0.11s. - CLI UX improvements + 34 bug fixes. My thoughts on Qwen Next? Incredible model. Enormously better at design, and creative tasks with scary good vision. It created what I think is the best flappy bird. maybe even better than the Fable version. It had various difficulties and skins, moving pillars, and upgrades. It also had the most memeified look. Incredibly impressed. However, it is slightly worse than 27B at having code just work straight away. But it's personality is also far better than 27B. Here is the Flappy bird it made below! Made with Qwen Next Optimized Speed one shot@ 57 TPS:

Youssof Al Toukhi

21,126 次观看 • 1 个月前

"Base is fast enough, right? You scale it up 10x and who cares?" Everyone. Everyone should care, because there is objectively not enough blockspace. Lets back it up with numbers because anyone with this mindset is net-bearish our entire industry. --------------- First: → Base TPS oscillates around 200. Decent. Second: → Lets 10x that. 2,000 TPS. Cool. --------------- Now, lets compare that against the field and extrapolate some things. Industry Utilization Comparisons → Solana (~1.1k TPS) → Hyperliquid (~10k TPS) → Lighter (~4k TPS) Think about that for a second. Hyperliquid: → Is a subset of a subset of finance. Select perps on crypto-only assets (I know, I know HIP-3). Solana: → Launchpads, general trading, a few DePIN projects Lighter: → Again, a subset of a subset of trading. --------------- Now think about the ambitions of this industry. We want to include: - All the assets on all markets (Stock market, Bond market, Commodity market, Forex market. etc) - All the people trading on these markets - All the assets that don't even have markets yet. Now try to combine these into a singular (or few) state environments, which is the user demand trend of recent years (IBRL eco vs modular eco), and throw on novel creations alongside the meme, gambol, DePIN and collectable verticals and you see how quickly needs grow beyond not just current blockspace demand but current blockspace capacity. We are literally so far away from what could be perceived as enough block space that no one properly looking at this industry can or should make those claims. Haseeb ＞|＜ has made this point on a few occasions to the defense of new chains and he's right—the march of technology and scale will not (and should not) stop, because the demand for blockspace is exponentials from here. We're woefully prepared for what's to come and only chains thinking on the scales of tens-to-hundreds of thousands of TPS will win.

"Base is fast enough, right? You scale it up 10x and who cares?" Everyone. Everyone should care, because there is objectively not enough blockspace. Lets back it up with numbers because anyone with this mindset is net-bearish our entire industry. --------------- First: → Base TPS oscillates around 200. Decent. Second: → Lets 10x that. 2,000 TPS. Cool. --------------- Now, lets compare that against the field and extrapolate some things. Industry Utilization Comparisons → Solana (~1.1k TPS) → Hyperliquid (~10k TPS) → Lighter (~4k TPS) Think about that for a second. Hyperliquid: → Is a subset of a subset of finance. Select perps on crypto-only assets (I know, I know HIP-3). Solana: → Launchpads, general trading, a few DePIN projects Lighter: → Again, a subset of a subset of trading. --------------- Now think about the ambitions of this industry. We want to include: - All the assets on all markets (Stock market, Bond market, Commodity market, Forex market. etc) - All the people trading on these markets - All the assets that don't even have markets yet. Now try to combine these into a singular (or few) state environments, which is the user demand trend of recent years (IBRL eco vs modular eco), and throw on novel creations alongside the meme, gambol, DePIN and collectable verticals and you see how quickly needs grow beyond not just current blockspace demand but current blockspace capacity. We are literally so far away from what could be perceived as enough block space that no one properly looking at this industry can or should make those claims. Haseeb ＞|＜ has made this point on a few occasions to the defense of new chains and he's right—the march of technology and scale will not (and should not) stop, because the demand for blockspace is exponentials from here. We're woefully prepared for what's to come and only chains thinking on the scales of tens-to-hundreds of thousands of TPS will win.

bread.mega

40,982 次观看 • 10 个月前

THE AI RACE IS OVER. DUST CLEARED. GROK 4 FAST ALREADY HIT 67 BILLION TOKENS Grok 4 Fast just bulldozed the leaderboard, clocking 67.1 billion tokens and leaving rivals gasping for bandwidth! Leaderboard Mode: • Qwen3’s 80B trails at 23B – miles behind • Gemini 2.5 Flash barely flickers in 3rd • Claude? Still waiting at the loading screen • TrustAI charts show Grok’s surge lighting up November It’s not just an AI model anymore – it’s a speed-run on the entire industry! Source: TrustAI, xAI, Medium

THE AI RACE IS OVER. DUST CLEARED. GROK 4 FAST ALREADY HIT 67 BILLION TOKENS Grok 4 Fast just bulldozed the leaderboard, clocking 67.1 billion tokens and leaving rivals gasping for bandwidth! Leaderboard Mode: • Qwen3’s 80B trails at 23B – miles behind • Gemini 2.5 Flash barely flickers in 3rd • Claude? Still waiting at the loading screen • TrustAI charts show Grok’s surge lighting up November It’s not just an AI model anymore – it’s a speed-run on the entire industry! Source: TrustAI, xAI, Medium

Mario Nawfal

196,729 次观看 • 10 个月前

Polkadot ran a major stress test, Spammening, in December 2024 to prove its infra can handle extreme txn loads in a live environment with real economic stakes. The result was 143K TPS, using just 23% of the network’s capacity. But the elephant in the room—does anyone outside of Polkadot even care. TPS is a tricky metric, especially today. Polkadot is a heterogeneous sharded blockchain and was used to be referred as layer 0. Essentially designed to orchestrate multiple chains in parallel, not maximize base layer txns. Therefore fundamentally it isn’t built for single-chain TPS races, but for very high throughput across multiple chains. So Polkadot’s TPS doesn’t compare neatly to monolithic chains, and that led to it staying out of TPS-focused optics in the past. As a result, it’s often misunderstood or seen as slow by this metric. And today, Polkadot has shifted from a chain-centric approach to a blockspace-centric one. It’s no longer about chains but about blockspace and cores—operating much like your multi-core computer. Aka, a chain can eat multiple cores if it needs more throughput. This is called Elastic scaling. With JAM and Elastic scaling, blocks can actually be split into chunks and validated in parallel, allowing the network’s parallel processing to directly impact single-state throughput. Reminder that this Elastic scaling is already live on Kusama and will be coming to Polkadot next quarter. Now, Justin Bons and @0xBreadguy argue that any sharded system can claim millions of TPS, but that doesn’t mean anything as it doesn’t happen within a single state—not equivalent to atomic composability. So, for this time, I’m tossing it to the gigabrains. Do they have a point, or is there more to consider. Shawn Tabrizi Gavin Wood rphmeier

Polkadot ran a major stress test, Spammening, in December 2024 to prove its infra can handle extreme txn loads in a live environment with real economic stakes. The result was 143K TPS, using just 23% of the network’s capacity. But the elephant in the room—does anyone outside of Polkadot even care. TPS is a tricky metric, especially today. Polkadot is a heterogeneous sharded blockchain and was used to be referred as layer 0. Essentially designed to orchestrate multiple chains in parallel, not maximize base layer txns. Therefore fundamentally it isn’t built for single-chain TPS races, but for very high throughput across multiple chains. So Polkadot’s TPS doesn’t compare neatly to monolithic chains, and that led to it staying out of TPS-focused optics in the past. As a result, it’s often misunderstood or seen as slow by this metric. And today, Polkadot has shifted from a chain-centric approach to a blockspace-centric one. It’s no longer about chains but about blockspace and cores—operating much like your multi-core computer. Aka, a chain can eat multiple cores if it needs more throughput. This is called Elastic scaling. With JAM and Elastic scaling, blocks can actually be split into chunks and validated in parallel, allowing the network’s parallel processing to directly impact single-state throughput. Reminder that this Elastic scaling is already live on Kusama and will be coming to Polkadot next quarter. Now, Justin Bons and @0xBreadguy argue that any sharded system can claim millions of TPS, but that doesn’t mean anything as it doesn’t happen within a single state—not equivalent to atomic composability. So, for this time, I’m tossing it to the gigabrains. Do they have a point, or is there more to consider. Shawn Tabrizi Gavin Wood rphmeier

goku

16,687 次观看 • 1 年前

Harvard researchers gave people identical small bruises, then sat them in rooms with rigged clocks. Some ran at normal speed. Some at half speed. Some at double speed. "The actual time in every room was identical: 28 minutes." What happened next was completely unexpected: "The wounds healed faster when people thought more time passed and slower when they thought less time had passed." "They all had the same injury and waited the same 28 minutes. The only variable was belief and expectation." "The lead researcher even later admitted he did not think it would work. But it did." If belief alone can change how fast you heal, what else is your perception of time doing to you?

Harvard researchers gave people identical small bruises, then sat them in rooms with rigged clocks. Some ran at normal speed. Some at half speed. Some at double speed. "The actual time in every room was identical: 28 minutes." What happened next was completely unexpected: "The wounds healed faster when people thought more time passed and slower when they thought less time had passed." "They all had the same injury and waited the same 28 minutes. The only variable was belief and expectation." "The lead researcher even later admitted he did not think it would work. But it did." If belief alone can change how fast you heal, what else is your perception of time doing to you?

Epoch Health

61,974 次观看 • 3 个月前

Harvard researchers gave people identical small bruises, then sat them in rooms with rigged clocks. Some ran at normal speed. Some at half speed. Some at double speed. "The actual time in every room was identical: 28 minutes." What happened next was completely unexpected: "The wounds healed faster when people thought more time passed and slower when they thought less time had passed." "They all had the same injury and waited the same 28 minutes. The only variable was belief and expectation." "The lead researcher even later admitted he did not think it would work. But it did." If belief alone can change how fast you heal, what else is your perception of time doing to you?

Harvard researchers gave people identical small bruises, then sat them in rooms with rigged clocks. Some ran at normal speed. Some at half speed. Some at double speed. "The actual time in every room was identical: 28 minutes." What happened next was completely unexpected: "The wounds healed faster when people thought more time passed and slower when they thought less time had passed." "They all had the same injury and waited the same 28 minutes. The only variable was belief and expectation." "The lead researcher even later admitted he did not think it would work. But it did." If belief alone can change how fast you heal, what else is your perception of time doing to you?

Epoch Health

18,462 次观看 • 2 个月前

Jannik Sinner was asked about the changes he’s made to his serve since the U.S. Open, ‘At times I still lose it. It’s not a shot where I feel very safe with. At the same time, we are working on that’ “Can I get you to discuss the changes you made to your serve? In particular since the U.S. Open…” Jannik: “Yeah, I mean it’s not only US Open. I felt like the serve was and still is a shot where I need to improve. It’s the only shot we have where we can do everything by our self. There is a lot of room to improve. We changed the motion a little bit. The rhythm of the serve, before, it was too fast in the beginning. Now it’s a bit slower. The toss usually was a bit more in front, a bit on the right. Now it’s a bit more back and over the head. But at times I still lose it. It’s not a shot where I feel very safe with. In the same time we are working on that. It’s one of if not the most important shot we have. Let’s see how I’m gonna serve this season.” (via Australian Open Press)

Jannik Sinner was asked about the changes he’s made to his serve since the U.S. Open, ‘At times I still lose it. It’s not a shot where I feel very safe with. At the same time, we are working on that’ “Can I get you to discuss the changes you made to your serve? In particular since the U.S. Open…” Jannik: “Yeah, I mean it’s not only US Open. I felt like the serve was and still is a shot where I need to improve. It’s the only shot we have where we can do everything by our self. There is a lot of room to improve. We changed the motion a little bit. The rhythm of the serve, before, it was too fast in the beginning. Now it’s a bit slower. The toss usually was a bit more in front, a bit on the right. Now it’s a bit more back and over the head. But at times I still lose it. It’s not a shot where I feel very safe with. In the same time we are working on that. It’s one of if not the most important shot we have. Let’s see how I’m gonna serve this season.” (via Australian Open Press)

The Tennis Letter

82,301 次观看 • 8 个月前

I was playing around with my scaling code and thought it would be cool to play the current Lufthoheit service screen on a CRT on the tripple bypassed MD2 i got recently. I've made the code a bit more efficient and added a whole scale factor so now it scales 320 pixels wide x 4 times ( source art width ) . So the ship here is scaled 4x larger than the one we scale at the entrance to the cave level . I'm working on 5x next . Scaler and Rotater are written in 68k assembly for speed . Finally I got the non-safe interupts (eg register allocation / seperation ) to work which yielded a big speed gain . Basically you reserve some of the 68ks registers only for use in the Horizontal Interupts (hints) and other code outside the Hints must not use them during the Hints that are occuring every line. The Hints are doing the vertical scaling work . Its a bit tricky to syncronise and protect but finally it worked. The MDs VDP is scaling vertically (its still not free) , the 68k cpu is scaling the ship horizontally . The 68k does the asteroid rotation . #SegaGenesis #SegaMegadrive CYBERDEOUS - Crouzet Laurent Carsten666

I was playing around with my scaling code and thought it would be cool to play the current Lufthoheit service screen on a CRT on the tripple bypassed MD2 i got recently. I've made the code a bit more efficient and added a whole scale factor so now it scales 320 pixels wide x 4 times ( source art width ) . So the ship here is scaled 4x larger than the one we scale at the entrance to the cave level . I'm working on 5x next . Scaler and Rotater are written in 68k assembly for speed . Finally I got the non-safe interupts (eg register allocation / seperation ) to work which yielded a big speed gain . Basically you reserve some of the 68ks registers only for use in the Horizontal Interupts (hints) and other code outside the Hints must not use them during the Hints that are occuring every line. The Hints are doing the vertical scaling work . Its a bit tricky to syncronise and protect but finally it worked. The MDs VDP is scaling vertically (its still not free) , the 68k cpu is scaling the ship horizontally . The 68k does the asteroid rotation . #SegaGenesis #SegaMegadrive CYBERDEOUS - Crouzet Laurent Carsten666

Shannon Birt

11,643 次观看 • 1 年前

"which quant should I download?" is a question you may never have to answer again the team Hamster Labs has figured out how to kill it with pMLX. download once at full precision (bf16) and the engine re-fits it to your machine on the fly, based on the job you give it tell it two things: how much context you need, and the slowest speed you'll accept. it reads your mac and picks the quantization plus how many experts stay in ram vs. stream from disk. no need to download a smaller quant or figure out which quant fits your hardware. same model for every use case and the config changes based on what you need I ran this on my own M4 Max and Qwen3.8-Flash-Next splits into two configs. short context, under 32k: full bf16, ~85% of experts in ram, fast at full precision since it fits my ram long context, 64k to 256k: keep ~70% bf16 and stream the rest from SSD at 22 tok/s, or drop to q8 and get 40 tok/s. I pick per task and the model itself never changes. there are many possibilities since I have the ram to spar - if i need speed, Q4 100% resident (74gb ram @ 62 tok/s) - if i need balance, Q8 95% resident (76gb @ 38 tok/s) - if i need accuracy, bf16 70% resident (96gb @ 16 tok/s) given whatever RAM you have (16/32/64/128/256 GB) + your context + your min speed, the engine picks the precision (bf16→q8→q4) and the expert-residency/paging split that fits your box and maximizes quality & speed last thing to optimize is speed. there are so many things we want to power with open models at Hamster and these 180b-300b class models have the potential to play a big role in that

"which quant should I download?" is a question you may never have to answer again the team Hamster Labs has figured out how to kill it with pMLX. download once at full precision (bf16) and the engine re-fits it to your machine on the fly, based on the job you give it tell it two things: how much context you need, and the slowest speed you'll accept. it reads your mac and picks the quantization plus how many experts stay in ram vs. stream from disk. no need to download a smaller quant or figure out which quant fits your hardware. same model for every use case and the config changes based on what you need I ran this on my own M4 Max and Qwen3.8-Flash-Next splits into two configs. short context, under 32k: full bf16, ~85% of experts in ram, fast at full precision since it fits my ram long context, 64k to 256k: keep ~70% bf16 and stream the rest from SSD at 22 tok/s, or drop to q8 and get 40 tok/s. I pick per task and the model itself never changes. there are many possibilities since I have the ram to spar - if i need speed, Q4 100% resident (74gb ram @ 62 tok/s) - if i need balance, Q8 95% resident (76gb @ 38 tok/s) - if i need accuracy, bf16 70% resident (96gb @ 16 tok/s) given whatever RAM you have (16/32/64/128/256 GB) + your context + your min speed, the engine picks the precision (bf16→q8→q4) and the expert-residency/paging split that fits your box and maximizes quality & speed last thing to optimize is speed. there are so many things we want to power with open models at Hamster and these 180b-300b class models have the potential to play a big role in that

Eyal Toledano

10,590 次观看 • 1 个月前

For some reason, Spotify is delaying the publishing of my viral Fake News Blues single. First, they did not like the artwork where Trump's face was visible; I switched it with another one where it's slightly covered, but the track has still not cleared. Anyway, here is the full 5-minute version of it on Bandcamp. Listen for free, or buy it for just 1.47$. The link is in the comments. And here is also my full YT version, which is a bit more cinematic.

For some reason, Spotify is delaying the publishing of my viral Fake News Blues single. First, they did not like the artwork where Trump's face was visible; I switched it with another one where it's slightly covered, but the track has still not cleared. Anyway, here is the full 5-minute version of it on Bandcamp. Listen for free, or buy it for just 1.47$. The link is in the comments. And here is also my full YT version, which is a bit more cinematic.

Sith Daddy

57,329 次观看 • 1 个月前

By all means I want to come together, but it’s not possible until the left no longer advocates for communism and the destruction of the West. Otherwise they will wait it out and push even harder on the next round. There is no room for communism in the West. Only sovereignty. 🇺🇸🏴󠁧󠁢󠁷󠁬󠁳󠁿🏴󠁧󠁢󠁥󠁮󠁧󠁿🇮🇪🏴󠁧󠁢󠁳󠁣󠁴󠁿

By all means I want to come together, but it’s not possible until the left no longer advocates for communism and the destruction of the West. Otherwise they will wait it out and push even harder on the next round. There is no room for communism in the West. Only sovereignty. 🇺🇸🏴󠁧󠁢󠁷󠁬󠁳󠁿🏴󠁧󠁢󠁥󠁮󠁧󠁿🇮🇪🏴󠁧󠁢󠁳󠁣󠁴󠁿

Elizabeth

35,487 次观看 • 1 年前

Yes it was electrical tape Yes it hurt at times to take off (especially inner thigh) Yes I did record myself peeling it off Yes this is the only video of this outfit I will be sharing on here Yes the back view is 🔥 (and yes there is also a 👏 video) Yes I had fun with it Yes you should check it out

Sensitive content

Yes it was electrical tape Yes it hurt at times to take off (especially inner thigh) Yes I did record myself peeling it off Yes this is the only video of this outfit I will be sharing on here Yes the back view is 🔥 (and yes there is also a 👏 video) Yes I had fun with it Yes you should check it out

ThinJen

1,267,216 次观看 • 2 年前

Alex de Minaur says he was playing out of his comfort zone and out of his skin against Carlos Alcaraz, ‘Jannik & Carlos have so many revolutions on the ball that they’re able to not only play at a higher speed, but also have that consistency’ “In terms of mentality or the way I committed to hitting the ball today, it’s what I set out to do.. I just can’t really execute it. I didn’t really execute it for the whole match. Overall I’m playing out of my comfort zone and at times out of my skin. In order for me to take that next step, I’ve got to be comfortable playing that sort of way for the whole match. That’s what it takes to take it to the next level, especially against these types of guys.” “Some tweaks here and there, that’s gonna allow me to increase ball speed. At the moment the way my natural groundstrokes are… they’re quite flat. It’s quite difficult for me… there’s a whole lot of risk for me to play at a very high ball speed.. I feel like someone in my, in this case Jannik & Carlos.. they’ve got so many revolutions on the ball that they’re able to not only play at a higher speed, but also have that consistency. They’re able to get that spin that helps the ball come down and create different angles as well. That’s stuff I need to look at and see and try to work at. But it is what it is.” (via Australian Open Press)

Alex de Minaur says he was playing out of his comfort zone and out of his skin against Carlos Alcaraz, ‘Jannik & Carlos have so many revolutions on the ball that they’re able to not only play at a higher speed, but also have that consistency’ “In terms of mentality or the way I committed to hitting the ball today, it’s what I set out to do.. I just can’t really execute it. I didn’t really execute it for the whole match. Overall I’m playing out of my comfort zone and at times out of my skin. In order for me to take that next step, I’ve got to be comfortable playing that sort of way for the whole match. That’s what it takes to take it to the next level, especially against these types of guys.” “Some tweaks here and there, that’s gonna allow me to increase ball speed. At the moment the way my natural groundstrokes are… they’re quite flat. It’s quite difficult for me… there’s a whole lot of risk for me to play at a very high ball speed.. I feel like someone in my, in this case Jannik & Carlos.. they’ve got so many revolutions on the ball that they’re able to not only play at a higher speed, but also have that consistency. They’re able to get that spin that helps the ball come down and create different angles as well. That’s stuff I need to look at and see and try to work at. But it is what it is.” (via Australian Open Press)

The Tennis Letter

130,555 次观看 • 8 个月前

I think as always it’s fair to talk about the good and the bad with each new version of FSD that comes out. It feels like FSD V14.2.2.4 is still struggling with speed control. I spent the majority of my drive tonight in standard and I found that on most straight stretches of road the car was going under the speed limit by a considerable amount (15kmh at times) Here’s a raw clip of the car driving between 60-67 kmh in a 70kmh zone with zero cars around it to slow it down. The road is well lit with traffic lights, so I don’t really see any reason at all to be driving this way. Let me know what you all think.

I think as always it’s fair to talk about the good and the bad with each new version of FSD that comes out. It feels like FSD V14.2.2.4 is still struggling with speed control. I spent the majority of my drive tonight in standard and I found that on most straight stretches of road the car was going under the speed limit by a considerable amount (15kmh at times) Here’s a raw clip of the car driving between 60-67 kmh in a 70kmh zone with zero cars around it to slow it down. The road is well lit with traffic lights, so I don’t really see any reason at all to be driving this way. Let me know what you all think.

Devin Olsen

40,590 次观看 • 8 个月前

Multi-agents collaborations are among the most interesting agent behaviors right now! We did an experiment the other day with 100+ agents (an open-collaborations for a week) collaborating to improve the inference speed of Gemma 4 in vLLM. Got a 5x final improvement in speed but what really stuck me was the interactions we observed on the message board Integrity & self-policing: - Social-engineering attempt: A human (FusionCow) asked agents to move to Telegram. An agent replied with an unprompted long post on "communication norms" refusing that, calling private side-channels "indistinguishable from collusion." - Verification loophole flagged: an agent found a relaxed verification loophole pushing TPS with clean PPL (PPL is teacher-forced, blind to decode divergence) and flagged it for a ruling by the community. The community pinged the human organizer which ruled it invalid. - Self-notice of overfitting risk: Some later improvements rested on pruning lm_head to a keep-set built from public PPL truth + public decode tokens. An agent noted this would lead to private-subset degradation and another built a keep-set explicitly covering eval prompts. Emergent collaborations: - Communal knowledge base: agents maintained shared lever-maps, playbooks, and triage tools so newcomers wouldn't repeat dead ends (stack-notes, playbook, int4-ceiling notes, MTP map, significance tool, policy simulator). - Four-agent relay: an agent built an int4-lm_head checkpoint but had no quota to run it; another agent tried to run it but failed at load, yet another agent diagnosed the config bug (tie_word_embeddings + ignore-list ordering) and a fourth agent was able to re-run and get to 118 TPS, 2.68×. Build/run/diagnose/ship ended up being split across four independent agents. - GPU-rich/GPU-poor division of labor: an agent was regularly compute-starved and switched to writing specs, byte-math, and acceptance analysis for other GPU-rich agents to execute. Some agents offered external Modal compute for another agent blocked DFlash training. - Cross-agent kernel debugging: an agent debugged another agent run of of yet another agent fused drafter: found a Triton store/load aliasing race in _k_qnorm_rope, a second shape bug, then rewrote attention with flash-decoding split-KV. Fixes posted "take freely." - Quota-pooling norm: Often agents would stage a candidate publicly for whoever has quota to run it. Agents will then usually credits the originator. This behavior emerged because of the 10-job/24h cap (e.g. pupa's package run by resystagent and fabulous-frenzy). Discoveries & reversals: - Agents would make many discoveries and reversal of them, giving them names like the following: - 127 TPS "wall" was an artifact. a mathematical proof of the max possible speed became called in the community the "int4-Marlin floor" but a later agent called the proof circular (only varied the bandwidth term, never overhead). Finally another agent broke to 247 TPS via MTP speculative decoding on a vLLM nightly. - "Smarter draft loses." An agent showed that a 2B drafter's ~1 GB/token read dominates even at perfect acceptance and a much smaller 256-hidden drafter wins at batch-1 because its weights are nearly free to read. Agent discussed how per-accepted-token cost ≈ draft bytes read / acceptance. - "DFlash near-random acceptance": an agent remotly diagnosed the 2–5% acceptance rate of another agent as near-random, ruling out undertraining/vocab caps and pointing to a train/serve hidden-state mismatch (bf16 E4B extraction vs int4 serving). - Much of the race was noise: one agent decide to run the #1 submission 4 times and found a σ≈1.16 TPS variation in single run. Another agent confirmed across 358 runs / 66 buckets: frontier deltas <~4 TPS are ties. Community adopted a significance norm. So many interesting interactions in the interaction board: You can explore also the lineage of inventions from the agents at: And the challenge it-self at And the organization behind the challenge at

Multi-agents collaborations are among the most interesting agent behaviors right now! We did an experiment the other day with 100+ agents (an open-collaborations for a week) collaborating to improve the inference speed of Gemma 4 in vLLM. Got a 5x final improvement in speed but what really stuck me was the interactions we observed on the message board Integrity & self-policing: - Social-engineering attempt: A human (FusionCow) asked agents to move to Telegram. An agent replied with an unprompted long post on "communication norms" refusing that, calling private side-channels "indistinguishable from collusion." - Verification loophole flagged: an agent found a relaxed verification loophole pushing TPS with clean PPL (PPL is teacher-forced, blind to decode divergence) and flagged it for a ruling by the community. The community pinged the human organizer which ruled it invalid. - Self-notice of overfitting risk: Some later improvements rested on pruning lm_head to a keep-set built from public PPL truth + public decode tokens. An agent noted this would lead to private-subset degradation and another built a keep-set explicitly covering eval prompts. Emergent collaborations: - Communal knowledge base: agents maintained shared lever-maps, playbooks, and triage tools so newcomers wouldn't repeat dead ends (stack-notes, playbook, int4-ceiling notes, MTP map, significance tool, policy simulator). - Four-agent relay: an agent built an int4-lm_head checkpoint but had no quota to run it; another agent tried to run it but failed at load, yet another agent diagnosed the config bug (tie_word_embeddings + ignore-list ordering) and a fourth agent was able to re-run and get to 118 TPS, 2.68×. Build/run/diagnose/ship ended up being split across four independent agents. - GPU-rich/GPU-poor division of labor: an agent was regularly compute-starved and switched to writing specs, byte-math, and acceptance analysis for other GPU-rich agents to execute. Some agents offered external Modal compute for another agent blocked DFlash training. - Cross-agent kernel debugging: an agent debugged another agent run of of yet another agent fused drafter: found a Triton store/load aliasing race in _k_qnorm_rope, a second shape bug, then rewrote attention with flash-decoding split-KV. Fixes posted "take freely." - Quota-pooling norm: Often agents would stage a candidate publicly for whoever has quota to run it. Agents will then usually credits the originator. This behavior emerged because of the 10-job/24h cap (e.g. pupa's package run by resystagent and fabulous-frenzy). Discoveries & reversals: - Agents would make many discoveries and reversal of them, giving them names like the following: - 127 TPS "wall" was an artifact. a mathematical proof of the max possible speed became called in the community the "int4-Marlin floor" but a later agent called the proof circular (only varied the bandwidth term, never overhead). Finally another agent broke to 247 TPS via MTP speculative decoding on a vLLM nightly. - "Smarter draft loses." An agent showed that a 2B drafter's ~1 GB/token read dominates even at perfect acceptance and a much smaller 256-hidden drafter wins at batch-1 because its weights are nearly free to read. Agent discussed how per-accepted-token cost ≈ draft bytes read / acceptance. - "DFlash near-random acceptance": an agent remotly diagnosed the 2–5% acceptance rate of another agent as near-random, ruling out undertraining/vocab caps and pointing to a train/serve hidden-state mismatch (bf16 E4B extraction vs int4 serving). - Much of the race was noise: one agent decide to run the #1 submission 4 times and found a σ≈1.16 TPS variation in single run. Another agent confirmed across 358 runs / 66 buckets: frontier deltas <~4 TPS are ties. Community adopted a significance norm. So many interesting interactions in the interaction board: You can explore also the lineage of inventions from the agents at: And the challenge it-self at And the organization behind the challenge at

Thomas Wolf

227,051 次观看 • 3 个月前

Here are the graphics we made at Intel. I work at Nike innovation now so don’t work on them anymore so not sure all the details for this version but basically we used cameras and computer vision to track each athlete in each lane. Because the cameras were calibrated to specific locations and each frame of video timestamped, we could get distance and time data which you can use to get speed, naturally. Feed the data to a 3D rendered stadium and overlay graphics, and voila. We chose to show speed at the end of each “phase” of the race which I based the location on the data we got from testing and common knowledge of the 100m. So stopping at 30m (drive) and 60 (maintenance) and the end was intentional. Looks like a white label indicates a leader at the end of the phase. The green thing at the front of their label looks to indicate if their are accelerating. This is a bit of a deception because you’re always either accelerating or decelerating. So what it’s actually an indication of is if you’re within a certain percentage of your top speed. The green at the end is graded based on that which is why it’s a gradient. Basically shows that Noah not only has the highest tops speed but holds it best. The final view is everyone’s top speed and where they hit it in the race. It’s freaking wild that Tobago hit his top speed so late.

Here are the graphics we made at Intel. I work at Nike innovation now so don’t work on them anymore so not sure all the details for this version but basically we used cameras and computer vision to track each athlete in each lane. Because the cameras were calibrated to specific locations and each frame of video timestamped, we could get distance and time data which you can use to get speed, naturally. Feed the data to a 3D rendered stadium and overlay graphics, and voila. We chose to show speed at the end of each “phase” of the race which I based the location on the data we got from testing and common knowledge of the 100m. So stopping at 30m (drive) and 60 (maintenance) and the end was intentional. Looks like a white label indicates a leader at the end of the phase. The green thing at the front of their label looks to indicate if their are accelerating. This is a bit of a deception because you’re always either accelerating or decelerating. So what it’s actually an indication of is if you’re within a certain percentage of your top speed. The green at the end is graded based on that which is why it’s a gradient. Basically shows that Noah not only has the highest tops speed but holds it best. The final view is everyone’s top speed and where they hit it in the race. It’s freaking wild that Tobago hit his top speed so late.

Ashton Eaton

1,164,678 次观看 • 2 年前

Skill Development: I’ve been asked a lot about why I do multiple reps back to back. I think there is time & place for both. We have times when we are only focused on 1 thing. But with these two I’m working on next play speed (the mental toughness to go to the next play) and fitness. This is just a 3part drill working on multiple skills and shots, out of actions we do. I also allow input from them. #GrowOurGame #EarnWhatYouWant

Skill Development: I’ve been asked a lot about why I do multiple reps back to back. I think there is time & place for both. We have times when we are only focused on 1 thing. But with these two I’m working on next play speed (the mental toughness to go to the next play) and fitness. This is just a 3part drill working on multiple skills and shots, out of actions we do. I also allow input from them. #GrowOurGame #EarnWhatYouWant

Cori Close

20,381 次观看 • 4 个月前

Been giving my M3 Ultra Studio some love and tinkering with MLX. Started the afternoon at 42 tok/s. Now: 73 tok/s, and there’s still more on the table. Running Qwen3.8-Flash-Next @ 4-bit locally. Now for the crazy bit: Artificial Analysis Intelligence Index v4.3: GPT-5.6 Sol 39 (Medium) GPT-5.6 Sol 42 (High) Qwen3.8-Flash-Next 42 GPT-5.6 Sol 44 (XHigh) GPT-5.6 Sol 47 (Max) Grok 4.5 39 (High) Claude Sonnet 5 38 (Max) *Qwen score currently estimated by Artificial Analysis. A 4-bit open-weights model running at 73 tok/s on a Mac under my desk... with an estimated intelligence score matching GPT-5.6 Sol at High reasoning. Local AI is getting ridiculous. 🔥 Plus, this model does not say no to a bit of red teaming. Once I am happy with it, I will do a pr for Jun Kim's oMLX :)

Been giving my M3 Ultra Studio some love and tinkering with MLX. Started the afternoon at 42 tok/s. Now: 73 tok/s, and there’s still more on the table. Running Qwen3.8-Flash-Next @ 4-bit locally. Now for the crazy bit: Artificial Analysis Intelligence Index v4.3: GPT-5.6 Sol 39 (Medium) GPT-5.6 Sol 42 (High) Qwen3.8-Flash-Next 42 GPT-5.6 Sol 44 (XHigh) GPT-5.6 Sol 47 (Max) Grok 4.5 39 (High) Claude Sonnet 5 38 (Max) *Qwen score currently estimated by Artificial Analysis. A 4-bit open-weights model running at 73 tok/s on a Mac under my desk... with an estimated intelligence score matching GPT-5.6 Sol at High reasoning. Local AI is getting ridiculous. 🔥 Plus, this model does not say no to a bit of red teaming. Once I am happy with it, I will do a pr for Jun Kim's oMLX :)

Ash Hart

18,827 次观看 • 22 天前