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。

SITUATION EXPLAINED: Cognition's new model is post-trained on Kimi K3. • SWE-2 scores 50.0% on FrontierCode 1.1 Main, Cognition's benchmark for whether a maintainer would merge the pull request • Fable 5.1 gets 50.9 at 64% higher cost. Grok 4.6 gets 48.0, Sol 47.5, Astra 53.3 • It leads... everything on Terminal-Bench 2.1 at 92.8 • Post-trained on Kimi K3. SWE-1.7 was post-trained on Kimi K2.7 • First Cognition model with effort levels, all from a single RL run • SWE-2 medium beats SWE-1.7 at 53 steps per run against 127, and 81% less cost Theo Jaffee: "Over the course of the RL run, they push the Pareto curve while preserving its shape. At every individual level of effort, from medium to high to max, it got better and also cheaper."show more

MTS

513,340 subscribers

13,488 次观看 • 10 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

1 条评论

Ricci Research 的头像

Ricci Research9 天前

53 steps versus 127 for the same result is the number with real economics behind it — agent costs scale with trajectory length, not parameter count, so halving the steps does more for unit economics than any pricing change. Also worth sitting with: an American coding company's frontier product is post-trained on Chinese open weights, and that's now just a routine architecture decision.

相关视频

We tested Kimi K3 against Claude Fable 5 on 3D modeling and animation Asked both to recreate the Millennium Falcon from Star Wars in 1) pixel style 2) original style. Both are on Max effort. Kimi K3 took a bit longer than Fable 5 to generate, but came in ~1/3 cheaper on the 1st test, and less than half the cost on the 2nd. Interestingly, Kimi K3 handled the details at the same level as Fable 5 on the original creation, yet it wasn't at Fable's level for the pixel style. Test 1 (Pixel) Kimi K3: 1hr 11min · $14.5 Fable 5: 49min · $21 Test 2 (Original) Kimi K3: 1hr 17min · $19.3 Fable 5: 1hr 5min · $47.2

We tested Kimi K3 against Claude Fable 5 on 3D modeling and animation Asked both to recreate the Millennium Falcon from Star Wars in 1) pixel style 2) original style. Both are on Max effort. Kimi K3 took a bit longer than Fable 5 to generate, but came in ~1/3 cheaper on the 1st test, and less than half the cost on the 2nd. Interestingly, Kimi K3 handled the details at the same level as Fable 5 on the original creation, yet it wasn't at Fable's level for the pixel style. Test 1 (Pixel) Kimi K3: 1hr 11min · $14.5 Fable 5: 49min · $21 Test 2 (Original) Kimi K3: 1hr 17min · $19.3 Fable 5: 1hr 5min · $47.2

GMI Cloud

80,618 次观看 • 2 个月前

$SITUATION EXPLAINED: Kimi K3 just beat Fable on a major coding benchmark, at a fraction of the price. • Kimi K3 beat Fable on Frontend Code Arena • The gap between Kimi and Fable is bigger than the gap between Fable and Sol • It also smashes Opus outright, while being priced the same as Sonnet, $3/$15 per million tokens • Sonnet 5 High currently sits in 10th place on this leaderboard at that same price point • The generational jump is dramatic: Kimi K2.6 ranked 18th on this benchmark, K3 jumped 17 spots to take the top spot • K3 ranked #1 in 6 of 7 tested domains: brand and marketing, reference-based design, data and analytics, consumer products, simulations, and content creation tools, it placed #2, only in gaming, behind Fable • Full model weights are set to release July 27th Theo Jaffee: "Imagine if Anthropic were a public company today, their stock would be down 20%, maybe more, enterprises are probably not gonna be spending $10 and $50 on Fable when they can instead be spending $3 and $15 on Kimi K3."$

SITUATION EXPLAINED: Kimi K3 just beat Fable on a major coding benchmark, at a fraction of the price. • Kimi K3 beat Fable on Frontend Code Arena • The gap between Kimi and Fable is bigger than the gap between Fable and Sol • It also smashes Opus outright, while being priced the same as Sonnet, $3/$15 per million tokens • Sonnet 5 High currently sits in 10th place on this leaderboard at that same price point • The generational jump is dramatic: Kimi K2.6 ranked 18th on this benchmark, K3 jumped 17 spots to take the top spot • K3 ranked #1 in 6 of 7 tested domains: brand and marketing, reference-based design, data and analytics, consumer products, simulations, and content creation tools, it placed #2, only in gaming, behind Fable • Full model weights are set to release July 27th Theo Jaffee: "Imagine if Anthropic were a public company today, their stock would be down 20%, maybe more, enterprises are probably not gonna be spending $10 and $50 on Fable when they can instead be spending $3 and $15 on Kimi K3."

MTS

132,930 次观看 • 2 个月前

Found a way to run Kimi K3 for FREE with daily credits on 😳 it's a 2.8T parameter model. 1M context. beats Fable 5 on coding benchmarks. what you get for $0: -2.8 trillion parameters -1M context window -beats Fable 5 on SWE Marathon (42 vs 35) -beats Fable 5 on Program Bench (77.8) -topped BrowseComp (91.2) -native vision understanding What this replaces: -Paid frontier API costs for coding agents -Fable 5 at $30/M output tokens -GPT-5.6 Sol at $25/M output tokens how to use it (2 min): 1. Go to 2. Sign up with email 3. Start using K3 — it's the default model Also available on desktop app, CLI, and mobile. Important: Free daily credits reset. Full open weights dropping July 27. Rate limits apply on free tier. use this before the free tier gets crowded

Found a way to run Kimi K3 for FREE with daily credits on 😳 it's a 2.8T parameter model. 1M context. beats Fable 5 on coding benchmarks. what you get for $0: -2.8 trillion parameters -1M context window -beats Fable 5 on SWE Marathon (42 vs 35) -beats Fable 5 on Program Bench (77.8) -topped BrowseComp (91.2) -native vision understanding What this replaces: -Paid frontier API costs for coding agents -Fable 5 at $30/M output tokens -GPT-5.6 Sol at $25/M output tokens how to use it (2 min): 1. Go to 2. Sign up with email 3. Start using K3 — it's the default model Also available on desktop app, CLI, and mobile. Important: Free daily credits reset. Full open weights dropping July 27. Rate limits apply on free tier. use this before the free tier gets crowded

painn

256,645 次观看 • 2 个月前

I tested Fable 5.1 vs Kimi K3 vs GPT 5.6 Sol vs Gemini 3.8 Flash I capped Fable, Kimi & Sol at $2 and asked them to generate this beach scene. Fable 5.1: Looks good at first glance, but made many mistakes, like people sitting on the volleyball court. Kimi K3: It did well. There were no major bugs like in Fable 5.1, and everything looks good. Sol: Bad at 3D scenes as usual, but Astra will definitely fix that. Gemini 3.8 Flash: I ran it in a loop for 35+ mins and it cost $0.95. It didn’t understand what it was fixing. The output here is the enhanced version. IMO, not even Terra level. Gemini's benchmarks say Opus 5-level, but maybe not for spatial understanding. Gemini has other use cases. Which one did the best job?

I tested Fable 5.1 vs Kimi K3 vs GPT 5.6 Sol vs Gemini 3.8 Flash I capped Fable, Kimi & Sol at $2 and asked them to generate this beach scene. Fable 5.1: Looks good at first glance, but made many mistakes, like people sitting on the volleyball court. Kimi K3: It did well. There were no major bugs like in Fable 5.1, and everything looks good. Sol: Bad at 3D scenes as usual, but Astra will definitely fix that. Gemini 3.8 Flash: I ran it in a loop for 35+ mins and it cost $0.95. It didn’t understand what it was fixing. The output here is the enhanced version. IMO, not even Terra level. Gemini's benchmarks say Opus 5-level, but maybe not for spatial understanding. Gemini has other use cases. Which one did the best job?

Bhavy☄️

33,971 次观看 • 17 天前

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.

Harman

41,278 次观看 • 2 个月前

Anthropic's in trouble, again. The entire Claude experience is now available at 1/6th the price. Kimi now does everything Claude does, powered by K2.6, a 1-trillion-parameter MoE model that activates only 32B parameters per token. It covers all three features Claude has (Chat, Code, and Cowork): 1) Kimi Chat runs in four modes - Instant for fast responses - Thinking for deep reasoning - Agent for multi-step execution - and Agent Swarm for parallel workloads. There's a 262K context window across all of them. 2) Kimi Code is the open-source CLI coding agent with K2.6 as the default backend. K2.6 ranked #1 on OpenRouter's programming leaderboard by weekly usage. 3) Kimi Agent is the Cowork equivalent. It generates: - full websites with database and auth - presentation decks (editable PPTX output) - spreadsheets with formulas and charts - word docs and structured research reports. On top of this, Kimi K2.6 is also trained to decompose tasks into up to 300 parallel sub-agents. This helps it retain coherence even across 4,000+ tool calls in a single run, with sessions sustaining up to 13 hours. On SWE-Bench Pro: - Kimi K2.6 → 58.6 - GPT-5.4 xhigh → 57.7 - Gemini 3.1 Pro → 54.2 - Claude Opus 4.6 → 53.4 Kimi K2.6 model is open weights and self-hostable on 4x H100s in INT4. Find the link to the HuggingFace model page in the replies!

Anthropic's in trouble, again. The entire Claude experience is now available at 1/6th the price. Kimi now does everything Claude does, powered by K2.6, a 1-trillion-parameter MoE model that activates only 32B parameters per token. It covers all three features Claude has (Chat, Code, and Cowork): 1) Kimi Chat runs in four modes - Instant for fast responses - Thinking for deep reasoning - Agent for multi-step execution - and Agent Swarm for parallel workloads. There's a 262K context window across all of them. 2) Kimi Code is the open-source CLI coding agent with K2.6 as the default backend. K2.6 ranked #1 on OpenRouter's programming leaderboard by weekly usage. 3) Kimi Agent is the Cowork equivalent. It generates: - full websites with database and auth - presentation decks (editable PPTX output) - spreadsheets with formulas and charts - word docs and structured research reports. On top of this, Kimi K2.6 is also trained to decompose tasks into up to 300 parallel sub-agents. This helps it retain coherence even across 4,000+ tool calls in a single run, with sessions sustaining up to 13 hours. On SWE-Bench Pro: - Kimi K2.6 → 58.6 - GPT-5.4 xhigh → 57.7 - Gemini 3.1 Pro → 54.2 - Claude Opus 4.6 → 53.4 Kimi K2.6 model is open weights and self-hostable on 4x H100s in INT4. Find the link to the HuggingFace model page in the replies!

Avi Chawla

110,057 次观看 • 4 个月前

$50 IN FREE CREDITS FOR GLM 5.2, KIMI K3 AND DEEPSEEK V4 PRO, ON A CODE THE COMPANY PUBLISHED ITSELF • ai&, the inference layer > Frontier models at up to 80% lower cost: > GLM 5.2, Kimi K3, Kimi K2.7 Code, DeepSeek V4 Pro and V4 Flash, plus Qwen3.6 27B, Gemma 4 31B and gpt-oss-120b. > OpenAI-compatible endpoint, so it drops into any agent or IDE you already run. • where the discount comes from > They route across AMD, NVIDIA and Tenstorrent, sending each workload to the silicon that suits it. > Everyone else tunes on one kind of chip, so the saving is architectural rather than promotional. Claiming it -> Credits in the top right, pick $50, apply "UNITABETAI" at checkout. The code has been live since their June launch, so run it on your own account before you plan around it ↓

$50 IN FREE CREDITS FOR GLM 5.2, KIMI K3 AND DEEPSEEK V4 PRO, ON A CODE THE COMPANY PUBLISHED ITSELF • ai&, the inference layer > Frontier models at up to 80% lower cost: > GLM 5.2, Kimi K3, Kimi K2.7 Code, DeepSeek V4 Pro and V4 Flash, plus Qwen3.6 27B, Gemma 4 31B and gpt-oss-120b. > OpenAI-compatible endpoint, so it drops into any agent or IDE you already run. • where the discount comes from > They route across AMD, NVIDIA and Tenstorrent, sending each workload to the silicon that suits it. > Everyone else tunes on one kind of chip, so the saving is architectural rather than promotional. Claiming it -> Credits in the top right, pick $50, apply "UNITABETAI" at checkout. The code has been live since their June launch, so run it on your own account before you plan around it ↓

slash1s

27,312 次观看 • 1 个月前

kimi k3 vs gpt 5.6 sol vs fable 5 vs grok 4.5 Kimi.ai just dropped kimi k3 – a 2.8t param native multimodal model, the first open 3t-class release. key facts: • 1m token context. stable latentmoe activating 16 of 896 experts, built on kimi delta attention (kda) and attention residuals • quantization-aware training from the sft stage onward – mxfp4 weights, mxfp8 activations. moonshot claims ~2.5x scaling efficiency over k2 • max thinking effort by default. low- and high-effort modes are "coming in updates" – there is no way to turn the thinking down today, and you feel it in every run • pricing: $0.30/mtok cache-hit input, $3.00/mtok cache-miss, $15.00/mtok output. claims >90% cache hit rate on coding workloads • benchmarks: swe marathon 42.0 (1st – fable 5: 35.0, sol: 39.0, opus 4.8: 40.0), terminal bench 2.1 88.3, browsecomp 91.2 (1st), program bench 77.8 (1st), gpqa-diamond 93.5. loses frontierswe 81.2 vs fable's 86.6, and deepswe 67.5 vs sol's 73.0 our test – 3 prompts, single-file html, Three.js, fully procedural, no assets: 1. photorealistic european roulette wheel – 37 pockets in the real sequence, mahogany clearcoat bowl, chrome turret, diamond deflectors, flick-to-spin, ball that spirals inward and settles on a mathematically real number 2. las vegas slot machine – 3 reels behind transmissive glass, drag the chrome lever to play, mechanical odometer counters modelled in 3d, coin physics on win 3. full pinball table – 6.5° tilted playfield, flipper impulse physics, spline ramps, drop targets, 6 bumpers, mechanical score reels in the backbox we ran the test on AI/ML API platform results: - cost #1 grok 4.5 – $0.30 #2 kimi k3 – $0.71 #3 gpt 5.6 sol – $2.05 #4 fable 5 – $7.69 - tokens #1 grok 4.5 – 34,241 #2 gpt 5.6 sol – 51,748 #3 fable 5 – 144,126 #4 kimi k3 – 157,999 - lines of code #1 gpt 5.6 sol – 3,054 #2 grok 4.5 – 3,047 #3 kimi k3 – 2,255 #4 fable 5 – 1,950 - generation time #1 grok 4.5 – 5.1 min #2 gpt 5.6 sol – 22.0 min #3 fable 5 – 31.5 min #4 kimi k3 – 75.6 min observations: • kimi k3 is cheap and it is slow. 75.6 minutes across three prompts against grok's 5.1. it is 2.4x grok's price and 15x grok's wall clock. the roulette took 15 min, the slot 18, the pinball 42 • it failed 2 of 3. only the roulette works. the slot machine has reel cutouts on both faces of the cabinet and the symbols face backwards – you can only read your spin by walking around to the rear of the machine. the pinball table stands vertically on its edge with the legs floating detached beside it. • 81% of kimi's output tokens are reasoning, not code. grok: 22%. you are not paying for a bigger answer, you are paying for a longer argument with itself • price per 100 shipped lines – grok $0.010, kimi $0.031, sol $0.067, fable $0.394. a 39x spread for the same three files kimi k3's code quality: upsides: • the roulette is genuinely good – procedural wood grain with real specular breakup, correct european sequence (0-32-15-19-4...), chrome turret, diamond deflectors, clean console • the pinball artwork is the best in the test – a synthwave "nova strike / deep space" field with six individually coloured neon bumper rings, a retro sun on a grid horizon, a nova burst, and a scoring legend printed on the apron. no other model printed the rules on the machine. it is a beautiful texture on a broken object • physics reasoning is real – it derived a 480hz substep for the collider, worked out ball settle conditions and termination guarantees, and checked every ramp exit vector by hand before writing any of it • it is the only model that saw the importmap trap coming. sol shipped a blank white page twice because three.js addons import the bare specifier 'three' and die without an import map downsides: • it dodged that trap on the slot by loading three.js r128 through classic script tags – a 2021 build with no working transmission. its slot glass rendered fully opaque and buried all three reels behind a white pane. the code asks for transmission: 0.93, ior: 1.5 – correct, and silently ignored by a renderer that predates the feature • after 42 minutes and 212k characters of reasoning, the pinball cabinet is not assembled. the table stands vertically on its edge like a wardrobe – the prompt asked for 6.5° from horizontal, it delivered 90°. the legs float detached in the void beside it. head-on it photographs beautifully; orbit ten degrees and it is a painted slab with four chrome rods hovering nearby • the playfield z-fights with the glass – hard black banding across the whole field as soon as you pull the camera back a note on the pinball, in fairness to kimi: nobody passed it. every model shipped broken ball physics and controls you cannot trust. it is the hardest prompt we have run and the whole field failed it, each in its own way kimi k3 reasons better than anything else here and it shows exactly where reasoning pays – physics constants, sequences, edge cases, traps the others walked into follow thehype. for 24/7 ai news, analysis and breakdowns

kimi k3 vs gpt 5.6 sol vs fable 5 vs grok 4.5 Kimi.ai just dropped kimi k3 – a 2.8t param native multimodal model, the first open 3t-class release. key facts: • 1m token context. stable latentmoe activating 16 of 896 experts, built on kimi delta attention (kda) and attention residuals • quantization-aware training from the sft stage onward – mxfp4 weights, mxfp8 activations. moonshot claims ~2.5x scaling efficiency over k2 • max thinking effort by default. low- and high-effort modes are "coming in updates" – there is no way to turn the thinking down today, and you feel it in every run • pricing: $0.30/mtok cache-hit input, $3.00/mtok cache-miss, $15.00/mtok output. claims >90% cache hit rate on coding workloads • benchmarks: swe marathon 42.0 (1st – fable 5: 35.0, sol: 39.0, opus 4.8: 40.0), terminal bench 2.1 88.3, browsecomp 91.2 (1st), program bench 77.8 (1st), gpqa-diamond 93.5. loses frontierswe 81.2 vs fable's 86.6, and deepswe 67.5 vs sol's 73.0 our test – 3 prompts, single-file html, Three.js, fully procedural, no assets: 1. photorealistic european roulette wheel – 37 pockets in the real sequence, mahogany clearcoat bowl, chrome turret, diamond deflectors, flick-to-spin, ball that spirals inward and settles on a mathematically real number 2. las vegas slot machine – 3 reels behind transmissive glass, drag the chrome lever to play, mechanical odometer counters modelled in 3d, coin physics on win 3. full pinball table – 6.5° tilted playfield, flipper impulse physics, spline ramps, drop targets, 6 bumpers, mechanical score reels in the backbox we ran the test on AI/ML API platform results: - cost #1 grok 4.5 – $0.30 #2 kimi k3 – $0.71 #3 gpt 5.6 sol – $2.05 #4 fable 5 – $7.69 - tokens #1 grok 4.5 – 34,241 #2 gpt 5.6 sol – 51,748 #3 fable 5 – 144,126 #4 kimi k3 – 157,999 - lines of code #1 gpt 5.6 sol – 3,054 #2 grok 4.5 – 3,047 #3 kimi k3 – 2,255 #4 fable 5 – 1,950 - generation time #1 grok 4.5 – 5.1 min #2 gpt 5.6 sol – 22.0 min #3 fable 5 – 31.5 min #4 kimi k3 – 75.6 min observations: • kimi k3 is cheap and it is slow. 75.6 minutes across three prompts against grok's 5.1. it is 2.4x grok's price and 15x grok's wall clock. the roulette took 15 min, the slot 18, the pinball 42 • it failed 2 of 3. only the roulette works. the slot machine has reel cutouts on both faces of the cabinet and the symbols face backwards – you can only read your spin by walking around to the rear of the machine. the pinball table stands vertically on its edge with the legs floating detached beside it. • 81% of kimi's output tokens are reasoning, not code. grok: 22%. you are not paying for a bigger answer, you are paying for a longer argument with itself • price per 100 shipped lines – grok $0.010, kimi $0.031, sol $0.067, fable $0.394. a 39x spread for the same three files kimi k3's code quality: upsides: • the roulette is genuinely good – procedural wood grain with real specular breakup, correct european sequence (0-32-15-19-4...), chrome turret, diamond deflectors, clean console • the pinball artwork is the best in the test – a synthwave "nova strike / deep space" field with six individually coloured neon bumper rings, a retro sun on a grid horizon, a nova burst, and a scoring legend printed on the apron. no other model printed the rules on the machine. it is a beautiful texture on a broken object • physics reasoning is real – it derived a 480hz substep for the collider, worked out ball settle conditions and termination guarantees, and checked every ramp exit vector by hand before writing any of it • it is the only model that saw the importmap trap coming. sol shipped a blank white page twice because three.js addons import the bare specifier 'three' and die without an import map downsides: • it dodged that trap on the slot by loading three.js r128 through classic script tags – a 2021 build with no working transmission. its slot glass rendered fully opaque and buried all three reels behind a white pane. the code asks for transmission: 0.93, ior: 1.5 – correct, and silently ignored by a renderer that predates the feature • after 42 minutes and 212k characters of reasoning, the pinball cabinet is not assembled. the table stands vertically on its edge like a wardrobe – the prompt asked for 6.5° from horizontal, it delivered 90°. the legs float detached in the void beside it. head-on it photographs beautifully; orbit ten degrees and it is a painted slab with four chrome rods hovering nearby • the playfield z-fights with the glass – hard black banding across the whole field as soon as you pull the camera back a note on the pinball, in fairness to kimi: nobody passed it. every model shipped broken ball physics and controls you cannot trust. it is the hardest prompt we have run and the whole field failed it, each in its own way kimi k3 reasons better than anything else here and it shows exactly where reasoning pays – physics constants, sequences, edge cases, traps the others walked into follow thehype. for 24/7 ai news, analysis and breakdowns

thehype.

2,187,777 次观看 • 2 个月前

My honest take on Kimi.ai K3 👀 As soon as I got access to Kimi, I put it through the exact same 3D football stadium challenge I had previously given Claude Fable 5. Fable 5 finished in under an hour. Kimi took almost 3 hours 🥵 At first, that was a very bad impression but i decided to see what's taking it this much of time. It was running E2E tests, validating desktop, tablet, and mobile, taking screenshots, finding failures, fixing them, even adapting for low end hardware and low FPS before moving on. All these that Fable 5 never did. None of it because it wasn't able to run this WebGL app on a headless browser. Kimi also shipped a clean React + Three.js codebase with proper components while Fable, as you might have checked codebase, generated one big HTML file with everything stuffed in. With all of this happening, it felt like I was doing more "Vibe Coding" with Fable 5 and more "Vibe Engineering" with Kimi K3. It feels like it's spending more time making sure the code actually works and easily scalable. Really impressed so far. More Kimi K3 experiments coming soon. Code (Kimi): Code (Fable): Live (Kimi hosted):

My honest take on Kimi.ai K3 👀 As soon as I got access to Kimi, I put it through the exact same 3D football stadium challenge I had previously given Claude Fable 5. Fable 5 finished in under an hour. Kimi took almost 3 hours 🥵 At first, that was a very bad impression but i decided to see what's taking it this much of time. It was running E2E tests, validating desktop, tablet, and mobile, taking screenshots, finding failures, fixing them, even adapting for low end hardware and low FPS before moving on. All these that Fable 5 never did. None of it because it wasn't able to run this WebGL app on a headless browser. Kimi also shipped a clean React + Three.js codebase with proper components while Fable, as you might have checked codebase, generated one big HTML file with everything stuffed in. With all of this happening, it felt like I was doing more "Vibe Coding" with Fable 5 and more "Vibe Engineering" with Kimi K3. It feels like it's spending more time making sure the code actually works and easily scalable. Really impressed so far. More Kimi K3 experiments coming soon. Code (Kimi): Code (Fable): Live (Kimi hosted):

The Bugged Dev

579,283 次观看 • 2 个月前

Open weight Kimi K3 performs at Claude Fable 5 level on 3D arcade games! We gave both models the same prompts: ✦ Pac-Man with four hunting ghosts ✦ Nokia Snake remade in 3D ✦ Retro space pinball table with real physics Outputs (thinking included): ✦ Kimi K3 (local, 8x B300): ~237K tokens, $0 ✦ Claude Fable 5: 92K tokens, $4.50 Every one of Kimi's tokens was free on our GPUs. Even via Moonshot's API the same Kimi run would be $3.60, still under Fable. Kimi held its quality against the strongest cloud model. Its snake is the liveliest scene of all six: soft shadows, floating dust, a flicking tongue, and a bot that checks it can still reach its own tail before every move. Fable's snake plays clean but looks flat next to it. Pac-Man is the closest call of the three. Kimi builds a fresh labyrinth on every restart and draws a live minimap in the corner while the red ghost tracks the player turn by turn through the corridors, and it even caught our bot on camera. Fable's maze answers with pearl pellets, +10 popups bursting over the floor and a wide-eyed cartoon ghost patrol. Fable won the pinball table: painted playfield art, a chrome ball that mirrors the neon lights, even little synth sound effects.

Open weight Kimi K3 performs at Claude Fable 5 level on 3D arcade games! We gave both models the same prompts: ✦ Pac-Man with four hunting ghosts ✦ Nokia Snake remade in 3D ✦ Retro space pinball table with real physics Outputs (thinking included): ✦ Kimi K3 (local, 8x B300): ~237K tokens, $0 ✦ Claude Fable 5: 92K tokens, $4.50 Every one of Kimi's tokens was free on our GPUs. Even via Moonshot's API the same Kimi run would be $3.60, still under Fable. Kimi held its quality against the strongest cloud model. Its snake is the liveliest scene of all six: soft shadows, floating dust, a flicking tongue, and a bot that checks it can still reach its own tail before every move. Fable's snake plays clean but looks flat next to it. Pac-Man is the closest call of the three. Kimi builds a fresh labyrinth on every restart and draws a live minimap in the corner while the red ghost tracks the player turn by turn through the corridors, and it even caught our bot on camera. Fable's maze answers with pearl pellets, +10 popups bursting over the floor and a wide-eyed cartoon ghost patrol. Fable won the pinball table: painted playfield art, a chrome ball that mirrors the neon lights, even little synth sound effects.

Atomic Agent

309,981 次观看 • 1 个月前

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

New Course: Post-training of LLMs Learn to post-train and customize an LLM in this short course, taught by Banghua Zhu, Assistant Professor at the University of Washington University of Washington, and co-founder of @NexusflowX. Training an LLM to follow instructions or answer questions has two key stages: pre-training and post-training. In pre-training, it learns to predict the next word or token from large amounts of unlabeled text. In post-training, it learns useful behaviors such as following instructions, tool use, and reasoning. Post-training transforms a general-purpose token predictor—trained on trillions of unlabeled text tokens—into an assistant that follows instructions and performs specific tasks. Because it is much cheaper than pre-training, it is practical for many more teams to incorporate post-training methods into their workflows than pre-training. In this course, you’ll learn three common post-training methods—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Online Reinforcement Learning (RL)—and how to use each one effectively. With SFT, you train the model on pairs of input and ideal output responses. With DPO, you provide both a preferred (chosen) and a less preferred (rejected) response and train the model to favor the preferred output. With RL, the model generates an output, receives a reward score based on human or automated feedback, and updates the model to improve performance. You’ll learn the basic concepts, common use cases, and principles for curating high-quality data for effective training. Through hands-on labs, you’ll download a pre-trained model from Hugging Face and post-train it using SFT, DPO, and RL to see how each technique shapes model behavior. In detail, you’ll: - Understand what post-training is, when to use it, and how it differs from pre-training. - Build an SFT pipeline to turn a base model into an instruct model. - Explore how DPO reshapes behavior by minimizing contrastive loss—penalizing poor responses and reinforcing preferred ones. - Implement a DPO pipeline to change the identity of a chat assistant. - Learn online RL methods such as Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), and how to design reward functions. - Train a model with GRPO to improve its math capabilities using a verifiable reward. Post-training is one of the most rapidly developing areas of LLM training. Whether you’re building a high-accuracy context-specific assistant, fine-tuning a model's tone, or improving task-specific accuracy, this course will give you experience with the most important techniques shaping how LLMs are post-trained today. Please sign up here:

Andrew Ng

125,146 次观看 • 1 年前

The truth about 5.6 sol after using it all day : The hype is overblown. Sort of. The benchmarks, and the commentary on X convinced me we were receiving AGI that runs at hyper speed, and is insanely cheap. in reality, 5.6 is built on the same spud pretraining as 5.5 this means its a nice bump, but not the opus to fable 5 LEAP in intelligence we recently experienced from anthropic. 5.6 is 2x times cheaper than fable on paper, and actually 3x cheaper when you look at actual task execution because of its token efficiency. BUT on swe bench where the models have to fix actual github bugs it falls behind fable pretty big. For vibecoders like myself this means I will be using 5.6 sol as a worker agent for Fable 5 to orchestrate alongside grok 4.5 I love this new era.

The truth about 5.6 sol after using it all day : The hype is overblown. Sort of. The benchmarks, and the commentary on X convinced me we were receiving AGI that runs at hyper speed, and is insanely cheap. in reality, 5.6 is built on the same spud pretraining as 5.5 this means its a nice bump, but not the opus to fable 5 LEAP in intelligence we recently experienced from anthropic. 5.6 is 2x times cheaper than fable on paper, and actually 3x cheaper when you look at actual task execution because of its token efficiency. BUT on swe bench where the models have to fix actual github bugs it falls behind fable pretty big. For vibecoders like myself this means I will be using 5.6 sol as a worker agent for Fable 5 to orchestrate alongside grok 4.5 I love this new era.

Max Blade

36,873 次观看 • 2 个月前

qwen 3.8 max vs deepseek v4 flash 0731 vs kimi k3 vs gpt 5.6 sol – on rubik's cube and chess four frontier models built a rubik's cube stand and solved it, then built a chess board and played claude opus 5 on it the setup: Nous Research's hermes agent cli on OpenRouter tasks: 1. cube – build a 3d rubik's cube with a cli and a Three.js viewer, then solve an identical scrambled position on your own stand 2. chess – build a 3d chess stand, then play white against claude opus 5 as black, live, one move at a time. no engine, no solver, no opening book on either side. stockfish depth 14 grades every chess ply afterwards; neither player sees the score models: DeepSeek v4 flash 0731, OpenAI gpt-5.6 sol, Kimi.ai kimi k3, Qwen qwen 3.8 max gpt-5.6 sol and deepseek v4 flash solved their cubes – sol in 24 moves and seventeen seconds, deepseek in 32. qwen and kimi never got there, giving up at 96 and 207 moves then all four built chess stands and played white against claude opus 5 on them, and all four resigned: deepseek on move 13, sol on 19, kimi on 21, qwen holding out longest at 29 - build time, both stands #1 gpt-5.6 sol – 16m 43s #2 deepseek v4 flash – 97m 39s #3 kimi k3 – 166m 09s #4 qwen 3.8 max – 215m 08s - build attempts before a working stand #1 gpt-5.6 sol – 3 #2 qwen 3.8 max – 4 #3 kimi k3 – 4 #4 deepseek v4 flash – 5 - total tokens #1 gpt-5.6 sol – 6,713,754 #2 qwen 3.8 max – 17,272,507 #3 kimi k3 – 22,427,504 #4 deepseek v4 flash – 27,417,442 - total price #1 deepseek v4 flash – $0.557 #2 gpt-5.6 sol – $6.319 #3 qwen 3.8 max – $10.270 #4 kimi k3 – $16.667 observations: • deepseek v4 flash is the cheapest model here by a margin nobody else is near, and it got there while being the least efficient of the four. it burned 27.4m tokens – more than anyone, 5m more than kimi – and still finished both benchmarks for $0.557. that is $0.02 per million tokens against kimi's $0.74. it also needed the most passes to produce working stands, five, and that did not matter: all five deepseek passes together cost a thirtieth of kimi's two • so what deepseek cannot do is get it right the first time. what it can do is get it right the fifth time, for half a dollar. that is a different thing to be buying – not a good first draft, but the option to keep asking • gpt-5.6 sol is the opposite profile and the strongest of the four on pure efficiency. 16m 43s to build both stands, 6.7m tokens, three passes – under 40% of the next lowest token count and a quarter of deepseek's, on an eighth of qwen's clock. it also solved the cube fastest of anyone, 24 moves in seventeen seconds. sol is what you reach for when you want the answer now and can absorb $0.94 per million • sol's weakness is in what it does not check. its chess viewer deleted the capturing piece instead of the captured one, so pieces disappeared off the board mid-game – a defect the fifty-cent deepseek stand did not have. fast and terse turns out to be the same dial as fast and unverified • qwen 3.8 max is not the cheap open-weights option it gets treated as. $10.270 across the two benchmarks, second most expensive of the four, 18x deepseek, and by a distance the slowest – 215 minutes of build time, nearly thirteen times sol's. what the money buys is judgment: it played eighteen moves without a single error worth a hundredth of a pawn, then made exactly one bad move in the whole game, and averaged 44.6 centipawns lost across the longest game any of the four managed. it also could not solve a rubik's cube in 96 tries • kimi k3 is the one line with no reading that flatters it. most expensive at $16.667, last on the cube at 207 moves, last at chess at 478 centipawns lost per move. it is also the model that verified hardest – on the cube it wrote its own integrity check instead of trusting its output. that makes the result worse rather than better: the checking was real, and the reasoning underneath it still was not follow thehype. for 24/7 ai news, analysis and breakdowns

thehype.

84,777 次观看 • 1 个月前

Chinese AI models are wiping billions off Big Tech right now. Google just lost $200 billion in a single day, and the model it needed to fight back still isn't ready. Gemini 3.5 Pro, Google's most powerful model, is months behind schedule. Alphabet stock dropped 4.4% that same day. The Deepseek moment is happening again, and the new model is FAR bigger. On the same day Google's delay leaked, a Beijing lab called Moonshot released Kimi K3. It is the largest open model ever built, with 2.8 trillion parameters. It took the number one spot on the Frontend Code Arena, a live coding leaderboard, passing Anthropic's best model. And Moonshot is giving it away for free on July 27. The genius part: Anyone with enough computers can download it and run a frontier level AI without paying a cent to a US company. A single task on Kimi K3 costs about 94 cents. The same work on some American models costs nearly double. So why would a company keep paying premium prices for a model it can now get for free? The entire US AI business is built on selling access to models that cost billions to train. If a free Chinese version does most of the same work, that pricing power starts to crack. And Kimi is close to the best. On one closely watched intelligence ranking it scored 57, just behind the top American models GPT-5.6 Sol and Fable 5, and ahead of Claude Opus 4.8. Bank of America told clients that Kimi proves Chinese labs can keep making big leaps even with limited chips. And the founder of Moonshot, Yang Zhilin, learned to build AI as a researcher INSIDE Google. Google literally wrote the 2017 paper that made all of these models possible. Now the people who studied its work are using it to destroy Google, and handing it out for free. What happens next: Kimi K3's weights go public on July 27. Google reports earnings on July 22, and everyone will be asking the same question about Gemini. If free models keep topping the charts, every valuation built on paid AI access has to be rewritten. What do you think?

Chinese AI models are wiping billions off Big Tech right now. Google just lost $200 billion in a single day, and the model it needed to fight back still isn't ready. Gemini 3.5 Pro, Google's most powerful model, is months behind schedule. Alphabet stock dropped 4.4% that same day. The Deepseek moment is happening again, and the new model is FAR bigger. On the same day Google's delay leaked, a Beijing lab called Moonshot released Kimi K3. It is the largest open model ever built, with 2.8 trillion parameters. It took the number one spot on the Frontend Code Arena, a live coding leaderboard, passing Anthropic's best model. And Moonshot is giving it away for free on July 27. The genius part: Anyone with enough computers can download it and run a frontier level AI without paying a cent to a US company. A single task on Kimi K3 costs about 94 cents. The same work on some American models costs nearly double. So why would a company keep paying premium prices for a model it can now get for free? The entire US AI business is built on selling access to models that cost billions to train. If a free Chinese version does most of the same work, that pricing power starts to crack. And Kimi is close to the best. On one closely watched intelligence ranking it scored 57, just behind the top American models GPT-5.6 Sol and Fable 5, and ahead of Claude Opus 4.8. Bank of America told clients that Kimi proves Chinese labs can keep making big leaps even with limited chips. And the founder of Moonshot, Yang Zhilin, learned to build AI as a researcher INSIDE Google. Google literally wrote the 2017 paper that made all of these models possible. Now the people who studied its work are using it to destroy Google, and handing it out for free. What happens next: Kimi K3's weights go public on July 27. Google reports earnings on July 22, and everyone will be asking the same question about Gemini. If free models keep topping the charts, every valuation built on paid AI access has to be rewritten. What do you think?

Ricardo

47,790 次观看 • 2 个月前

a screenshot from openai's internal slack leaked last night, and sam altman is telling his own staff to stop comparing gpt-6 astra to the free models "in writing." when the ceo bans the comparison, the comparison is already over. the leaked line is one sentence: do not benchmark astra against open weights in any channel. you do not ban a fight you are winning. altman just told 4,000 employees the free model caught up, and one of them told the internet. here is the comparison he did not want on the record, and it runs local for about $6 instead of $400: -> pull an open model like grok or kimi k3, the exact ones altman just barred his own team from benchmarking, and load it onto one nvidia card -> point cursor, cline and every app at localhost, and none of them notice they stopped paying for astra -> keep the heavy jobs on a rented gpu by the hour and the daily work fully offline for the cost of power -> your $200 chatgpt seat and your $200 claude seat both go dark, and no memo can put that back here is the part they will fight me on: openai and anthropic are not banning the benchmark because it is wrong, they are banning it because it is right. the day a lab tells its own people to stop measuring, the $200 stops buying a better model and starts buying the silence around a tie. drop your $400 stack to a $6 local model and run the comparison altman deleted from his own slack. the full benchmark is in the article below, save it before it gets "clarified."

a screenshot from openai's internal slack leaked last night, and sam altman is telling his own staff to stop comparing gpt-6 astra to the free models "in writing." when the ceo bans the comparison, the comparison is already over. the leaked line is one sentence: do not benchmark astra against open weights in any channel. you do not ban a fight you are winning. altman just told 4,000 employees the free model caught up, and one of them told the internet. here is the comparison he did not want on the record, and it runs local for about $6 instead of $400: -> pull an open model like grok or kimi k3, the exact ones altman just barred his own team from benchmarking, and load it onto one nvidia card -> point cursor, cline and every app at localhost, and none of them notice they stopped paying for astra -> keep the heavy jobs on a rented gpu by the hour and the daily work fully offline for the cost of power -> your $200 chatgpt seat and your $200 claude seat both go dark, and no memo can put that back here is the part they will fight me on: openai and anthropic are not banning the benchmark because it is wrong, they are banning it because it is right. the day a lab tells its own people to stop measuring, the $200 stops buying a better model and starts buying the silence around a tie. drop your $400 stack to a $6 local model and run the comparison altman deleted from his own slack. the full benchmark is in the article below, save it before it gets "clarified."

starmex

24,200 次观看 • 7 天前

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has *all* the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has all the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Sonya Huang 🐥

81,332 次观看 • 3 个月前

FREE CLAUDE FABLE 5 and almost nobody knows this trick Anthropic recently released their most capable model ever - the first Mythos-class model available to the public normally it's $10/$50 per million tokens but there's a hack: GitLab just added Fable 5 to Duo Agent Platform - across ALL tiers, including free trial the setup: 1/ register a GitLab account: 2/ start the free GitLab Duo trial 3/ Fable 5 is live through their AI Gateway 4/ done - Mythos-class model, $0 why Fable 5 is worth the hype: 1/ 80.3% SWE-Bench Pro - 11 points ahead of the next best model 2/ multi-DAY autonomous runs without losing the plot 3/ single-pass implementations of systems that took days of iteration 4/ Karpathy called it 'a major-version-bump-deserving step change' also free on Claude Pro/Max/Team until June 22 - after that it's usage credits you have a limited window to run the most capable public model on the planet for FREE

FREE CLAUDE FABLE 5 and almost nobody knows this trick Anthropic recently released their most capable model ever - the first Mythos-class model available to the public normally it's $10/$50 per million tokens but there's a hack: GitLab just added Fable 5 to Duo Agent Platform - across ALL tiers, including free trial the setup: 1/ register a GitLab account: 2/ start the free GitLab Duo trial 3/ Fable 5 is live through their AI Gateway 4/ done - Mythos-class model, $0 why Fable 5 is worth the hype: 1/ 80.3% SWE-Bench Pro - 11 points ahead of the next best model 2/ multi-DAY autonomous runs without losing the plot 3/ single-pass implementations of systems that took days of iteration 4/ Karpathy called it 'a major-version-bump-deserving step change' also free on Claude Pro/Max/Team until June 22 - after that it's usage credits you have a limited window to run the most capable public model on the planet for FREE

kaize

24,587 次观看 • 3 个月前

🚨 do you understand what just happened to Microsoft.. Microsoft just dropped seven of its own MAI models, trained from scratch with zero distillation, and said its custom-tuned models already match GPT-tier quality at 10x lower cost. The partner that pays OpenAI's bills is now quietly building the thing that replaces it. - MAI-Thinking-1 hits human-preference parity with Sonnet 4.6 and 53% on SWE Bench Pro, right next to Opus 4.6 - MAI-Code-1-Flash delivers Haiku-class coding at just 5B params and is already shipping inside GitHub Copilot - A custom MAI model tuned for Excel matched GPT-5.4 while running up to 10x more efficiently - Their real weapon isn't the weights, it's that they're already inside everyone's Office, Teams, and VS Code The hard part was never training a good model. The hard part is being the room everyone already works in. Microsoft owns the room.

🚨 do you understand what just happened to Microsoft.. Microsoft just dropped seven of its own MAI models, trained from scratch with zero distillation, and said its custom-tuned models already match GPT-tier quality at 10x lower cost. The partner that pays OpenAI's bills is now quietly building the thing that replaces it. - MAI-Thinking-1 hits human-preference parity with Sonnet 4.6 and 53% on SWE Bench Pro, right next to Opus 4.6 - MAI-Code-1-Flash delivers Haiku-class coding at just 5B params and is already shipping inside GitHub Copilot - A custom MAI model tuned for Excel matched GPT-5.4 while running up to 10x more efficiently - Their real weapon isn't the weights, it's that they're already inside everyone's Office, Teams, and VS Code The hard part was never training a good model. The hard part is being the room everyone already works in. Microsoft owns the room.

BuBBliK

171,865 次观看 • 3 个月前

BREAKING: OpenAI just dropped GPT-6 ASTRA!!! 🚀✨ We’ve been testing it extensively at Every 📧 across coding, writing, and knowledge work. My take: it’s a big upgrade from 5.6-Sol, with some frustrating habits that keep it from matching Fable at the top end. Here’s your vibe check: - The best writing model I’ve tried. It’s fast, produces very little slop, and is easy to steer. It’s a good companion for actually working through the writing I do every day. (Not to mention, it one-shotted the first draft of its own vibe check today!) - The computer use is wild. It can go for hours at a time using complicated apps to get work done. It did the first cut of our Fable 5.1 vibe check video...kind of mindblowing - Impressive 3D games and visualizations. It can make beautiful 3D worlds from a single prompt. I one-shotted a historically accurate rendition of the Battle of Waterloo - It can overcomplicate things. (Especially at higher effort levels.) Ask for a simple interface and you get extra labels, buttons, and features everywhere. It has a habit of turning everything into a landing page. It just doesnt quite match Fable's ability to intuitively understand your prompt and do something delightful (without overcomplicating.) Net Result: If you already live in ChatGPT for Work or Codex and can afford it, it’s an easy upgrade from 5.6-Sol. The biggest proof of Astra's effectiveness at helping you do work is our vibe check. We found out it was launching at 3 AM this morning, and had a 4,000 word vibe check + video done by 2 PM. Not possible without this model. I’m reaching for Astra all day, but Fable 5.1 still gets my biggest tasks. On ambitious builds, Fable is better at understanding what I want and taking it further than I would have thought to ask. State of Play: Astra is launching to Enterprise customers today, and the rest of ChatGPT users over the coming days. Now, both OpenAI and Anthropic have a higher class of models that cost more to use. That changes who gets to use frontier AI and how. It's also a new vector of competition between them: Fable and Astra are priced at the same level. We'll see what that means for adoption in the coming days and weeks. read our full vibe check Every 📧 today:

BREAKING: OpenAI just dropped GPT-6 ASTRA!!! 🚀✨ We’ve been testing it extensively at Every 📧 across coding, writing, and knowledge work. My take: it’s a big upgrade from 5.6-Sol, with some frustrating habits that keep it from matching Fable at the top end. Here’s your vibe check: - The best writing model I’ve tried. It’s fast, produces very little slop, and is easy to steer. It’s a good companion for actually working through the writing I do every day. (Not to mention, it one-shotted the first draft of its own vibe check today!) - The computer use is wild. It can go for hours at a time using complicated apps to get work done. It did the first cut of our Fable 5.1 vibe check video...kind of mindblowing - Impressive 3D games and visualizations. It can make beautiful 3D worlds from a single prompt. I one-shotted a historically accurate rendition of the Battle of Waterloo - It can overcomplicate things. (Especially at higher effort levels.) Ask for a simple interface and you get extra labels, buttons, and features everywhere. It has a habit of turning everything into a landing page. It just doesnt quite match Fable's ability to intuitively understand your prompt and do something delightful (without overcomplicating.) Net Result: If you already live in ChatGPT for Work or Codex and can afford it, it’s an easy upgrade from 5.6-Sol. The biggest proof of Astra's effectiveness at helping you do work is our vibe check. We found out it was launching at 3 AM this morning, and had a 4,000 word vibe check + video done by 2 PM. Not possible without this model. I’m reaching for Astra all day, but Fable 5.1 still gets my biggest tasks. On ambitious builds, Fable is better at understanding what I want and taking it further than I would have thought to ask. State of Play: Astra is launching to Enterprise customers today, and the rest of ChatGPT users over the coming days. Now, both OpenAI and Anthropic have a higher class of models that cost more to use. That changes who gets to use frontier AI and how. It's also a new vector of competition between them: Fable and Astra are priced at the same level. We'll see what that means for adoption in the coming days and weeks. read our full vibe check Every 📧 today:

Dan Shipper

387,231 次观看 • 17 天前

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

kwindla

40,593 次观看 • 3 个月前