also how the hell do you compare the prices of these things. like how many kilowatt hours of electricity for a genome? how do you even *measure* cost of AI given the differing performance-per-token numbers between models
October 2, 2026 at 8:47 PM
also how the hell do you compare the prices of these things. like how many kilowatt hours of electricity for a genome? how do you even *measure* cost of AI given the differing performance-per-token numbers between models
COPD, a contrastive on-policy distillation framework, enhances reasoning efficiency via comparative token evaluations. This method reduces response lengths while keeping accuracy across multimodal benchmarks, improving performance for complex problem-solving. https://arxiv.org/abs/2607.19046
Contrastive On-Policy Distillation
ArXiv link for Contrastive On-Policy Distillation
arxiv.org
October 2, 2026 at 7:00 PM
COPD, a contrastive on-policy distillation framework, enhances reasoning efficiency via comparative token evaluations. This method reduces response lengths while keeping accuracy across multimodal benchmarks, improving performance for complex problem-solving. https://arxiv.org/abs/2607.19046
Capcom RE:Dox, open source serialization/deserialization for game engines
Discussion | hackernews | Author: vblanco
#Performance
Discussion | hackernews | Author: vblanco
#Performance
Capcom RE:Dox, open source serialization/deserialization for game engines
High-performance, token-based structured data engine for .NET. A core component of REX, the technology behind CAPCOM's next-generation game engine. - CAPCOM-TD-OSS/REDox
github.com
October 2, 2026 at 6:36 PM
Capcom RE:Dox, open source serialization/deserialization for game engines
Discussion | hackernews | Author: vblanco
#Performance
Discussion | hackernews | Author: vblanco
#Performance
Somewhat tangential but quantization errors cancel and quantized models preserve the weight on the top-ranked token.
www.alphaxiv.org/abs/2609.11716
www.alphaxiv.org/abs/2609.11716
Why Does Post-Training Quantization Work?
Researchers investigated the underlying mechanisms that enable pretrained large language models to maintain performance under post-training quantization, identifying a "counteracting residual...
www.alphaxiv.org
October 2, 2026 at 6:17 PM
Somewhat tangential but quantization errors cancel and quantized models preserve the weight on the top-ranked token.
www.alphaxiv.org/abs/2609.11716
www.alphaxiv.org/abs/2609.11716
Google Gemini 4 Argon: Frontier Performance Meets Benchmaxxing Debate www.kad8.com/ai/google-ge...
Google Gemini 4 Argon: Frontier Performance Meets Benchmaxxing Debate
Google's Gemini 4 Argon targets frontier AI performance with million-token output and engineering gains, while benchmark optimization faces scrutiny.
www.kad8.com
October 2, 2026 at 5:59 PM
Google Gemini 4 Argon: Frontier Performance Meets Benchmaxxing Debate www.kad8.com/ai/google-ge...
this is a nice example of a class of thing that I like to bundle under "making tools more effective for AI is an excuse to make them better for humans" - Capcom made a super-fast text-backup serializer/deserializer for .NET to support better perf with text data formats github.com/CAPCOM-TD-OS...
GitHub - CAPCOM-TD-OSS/REDox: High-performance, token-based structured data engine for .NET. A core component of REX, the technology behind CAPCOM's next-generation game engine.
High-performance, token-based structured data engine for .NET. A core component of REX, the technology behind CAPCOM's next-generation game engine. - CAPCOM-TD-OSS/REDox
github.com
October 2, 2026 at 5:26 PM
this is a nice example of a class of thing that I like to bundle under "making tools more effective for AI is an excuse to make them better for humans" - Capcom made a super-fast text-backup serializer/deserializer for .NET to support better perf with text data formats github.com/CAPCOM-TD-OS...
Starting a rumor that Capcom made their own JSON library because the JSON license says it can’t be used “for evil” (www.json.org/license.html) and their lawyers determined that included “Resident Evil” github.com/CAPCOM-TD-OS...
GitHub - CAPCOM-TD-OSS/REDox: High-performance, token-based structured data engine for .NET. A core component of REX, the technology behind CAPCOM's next-generation game engine.
High-performance, token-based structured data engine for .NET. A core component of REX, the technology behind CAPCOM's next-generation game engine. - CAPCOM-TD-OSS/REDox
github.com
October 2, 2026 at 2:24 PM
Starting a rumor that Capcom made their own JSON library because the JSON license says it can’t be used “for evil” (www.json.org/license.html) and their lawyers determined that included “Resident Evil” github.com/CAPCOM-TD-OS...
OpenAI's GPT-6 Astra Ultrafast is live on API. NVIDIA Blackwell infrastructure delivers up to 8× faster token generation compared to previous deployments. This is the performance ceiling of the GPT-6 family — frontier-quality, near-real-time inference, now commercially available. The 8× number is...
How NVIDIA GPUs Help Accelerate OpenAI’s GPT-6 Astra Ultrafast
Running on NVIDIA Blackwell GPUs and accelerated by continuous inference optimizations through OpenAI’s models, GPT-6 Astra Ultrafast delivers faster model responses across code generation, tool use and interactive applications.
blogs.nvidia.com
October 2, 2026 at 7:00 AM
OpenAI's GPT-6 Astra Ultrafast is live on API. NVIDIA Blackwell infrastructure delivers up to 8× faster token generation compared to previous deployments. This is the performance ceiling of the GPT-6 family — frontier-quality, near-real-time inference, now commercially available. The 8× number is...
RE:Dox, high performance token-based structured data engine for .NET https://www.reddit.com/r/dotnet/comments/1wvlpgt/redox_high_performance_tokenbased_structured_data/
#dotnet
#dotnet
October 2, 2026 at 5:44 AM
RE:Dox, high performance token-based structured data engine for .NET https://www.reddit.com/r/dotnet/comments/1wvlpgt/redox_high_performance_tokenbased_structured_data/
#dotnet
#dotnet
Capcom has opensourced RE:Dox, a JSON data engine part of the REX project. Check it out here:
github.com/CAPCOM-TD-OS...
#opensource #dotnet
github.com/CAPCOM-TD-OS...
#opensource #dotnet
GitHub - CAPCOM-TD-OSS/REDox: High-performance, token-based structured data engine for .NET. A core component of REX, the technology behind CAPCOM's next-generation game engine.
High-performance, token-based structured data engine for .NET. A core component of REX, the technology behind CAPCOM's next-generation game engine. - CAPCOM-TD-OSS/REDox
github.com
October 2, 2026 at 5:23 AM
Capcom has opensourced RE:Dox, a JSON data engine part of the REX project. Check it out here:
github.com/CAPCOM-TD-OS...
#opensource #dotnet
github.com/CAPCOM-TD-OS...
#opensource #dotnet
traffic: AI performance reports token usage and model latency for the large language models called through your gateways, with filters for model and provider. Tool performance reports traffic, throughput, payload sizes, and latency for your Model Context Protocol (MCP)
October 2, 2026 at 4:20 AM
traffic: AI performance reports token usage and model latency for the large language models called through your gateways, with filters for model and provider. Tool performance reports traffic, throughput, payload sizes, and latency for your Model Context Protocol (MCP)
NAMOH is a sparse attention mechanism enhancing long-context processing in language models. Activating a subset of heads per token increases efficiency while maintaining performance, outpacing fully activated models and optimizing context retrieval for AI scaling. https://arxiv.org/abs/2609.38832
Scaling Parameter and Context in Attention: Native Sparse Attention from Mixture-of-Head
ArXiv link for Scaling Parameter and Context in Attention: Native Sparse Attention from Mixture-of-Head
arxiv.org
October 2, 2026 at 3:00 AM
NAMOH is a sparse attention mechanism enhancing long-context processing in language models. Activating a subset of heads per token increases efficiency while maintaining performance, outpacing fully activated models and optimizing context retrieval for AI scaling. https://arxiv.org/abs/2609.38832
📁 Sectors Performance (%):
・ 1. STX.CITY Ecosystem - $FLAT, $SKULL
・ 2. Flaunch Ecosystem - $FLNCH, $GITB
・ 3. Decentralized Social Media (DeSOC) - $MCAT, $DESO, $BSO
・ 4. ERC 404 - $PANDO, $DEFRO, $PURSE
・ 5. Printr Launchpad - $DEPLO, $发财, $OOO
・ 1. STX.CITY Ecosystem - $FLAT, $SKULL
・ 2. Flaunch Ecosystem - $FLNCH, $GITB
・ 3. Decentralized Social Media (DeSOC) - $MCAT, $DESO, $BSO
・ 4. ERC 404 - $PANDO, $DEFRO, $PURSE
・ 5. Printr Launchpad - $DEPLO, $发财, $OOO
STX.City — No-code tools for the Stacks ecosystem
Build, deploy, and monitor on Stacks without writing a line of Clarity. Token launchpad, bonding curve, BNS, analytics, and more — all in one dashboard.
STX.CITY
October 2, 2026 at 1:48 AM
📁 Sectors Performance (%):
・ 1. STX.CITY Ecosystem - $FLAT, $SKULL
・ 2. Flaunch Ecosystem - $FLNCH, $GITB
・ 3. Decentralized Social Media (DeSOC) - $MCAT, $DESO, $BSO
・ 4. ERC 404 - $PANDO, $DEFRO, $PURSE
・ 5. Printr Launchpad - $DEPLO, $发财, $OOO
・ 1. STX.CITY Ecosystem - $FLAT, $SKULL
・ 2. Flaunch Ecosystem - $FLNCH, $GITB
・ 3. Decentralized Social Media (DeSOC) - $MCAT, $DESO, $BSO
・ 4. ERC 404 - $PANDO, $DEFRO, $PURSE
・ 5. Printr Launchpad - $DEPLO, $发财, $OOO
NAMOH reveals a native sparse attention technique that activates heads per token, allowing for efficient long-context handling. This method improves performance while cutting down memory and computation needs, thus fostering scalable language models. https://arxiv.org/abs/2609.38832
Scaling Parameter and Context in Attention: Native Sparse Attention from Mixture-of-Head
ArXiv link for Scaling Parameter and Context in Attention: Native Sparse Attention from Mixture-of-Head
arxiv.org
October 1, 2026 at 8:40 PM
NAMOH reveals a native sparse attention technique that activates heads per token, allowing for efficient long-context handling. This method improves performance while cutting down memory and computation needs, thus fostering scalable language models. https://arxiv.org/abs/2609.38832
Tokenomics is more than counting tokens. The real question is whether AI spending delivers successful outcomes -- and whether the cost, performance and risk are in balance.
virtualizationreview.com/articles/202...
virtualizationreview.com/articles/202...
Why Tokenomics Is More Than Just Counting Tokens -- Virtualization Review
Tokenomics should measure more than token consumption -- organizations need to balance cost, performance, risk and business value.
virtualizationreview.com
October 1, 2026 at 7:12 PM
Tokenomics is more than counting tokens. The real question is whether AI spending delivers successful outcomes -- and whether the cost, performance and risk are in balance.
virtualizationreview.com/articles/202...
virtualizationreview.com/articles/202...
Cut your AI spend with AI Gateway's Auto Router
Cloudflare AI Gateway now features a model router that evaluates request complexity using an edge-deployed classifier to select the optimal model. By balancing expected output quality against token costs, organizations c…
Telegram AI Digest
#ai #news
Cloudflare AI Gateway now features a model router that evaluates request complexity using an edge-deployed classifier to select the optimal model. By balancing expected output quality against token costs, organizations c…
Telegram AI Digest
#ai #news
Cut your AI spend with AI Gateway's Auto Router
Cloudflare AI Gateway now features a model router that evaluates request complexity using an edge-deployed classifier to select the optimal model. By balancing expected output quality against token costs, organizations can dramatically cut AI spend while maintaining performance.
blog.cloudflare.com
October 1, 2026 at 4:14 PM
Cut your AI spend with AI Gateway's Auto Router
Cloudflare AI Gateway now features a model router that evaluates request complexity using an edge-deployed classifier to select the optimal model. By balancing expected output quality against token costs, organizations c…
Telegram AI Digest
#ai #news
Cloudflare AI Gateway now features a model router that evaluates request complexity using an edge-deployed classifier to select the optimal model. By balancing expected output quality against token costs, organizations c…
Telegram AI Digest
#ai #news
Google announced Gemini 4 Argon Sept. 30; rollout is limited to trusted testers and cyber defenders via Fairwind, with no public-release date. Its claimed 1M-token output could matter for coding and enterprise work, but benchmark claims and coding performance face questions. #AI #Cybersecurity
October 1, 2026 at 11:24 AM
Google announced Gemini 4 Argon Sept. 30; rollout is limited to trusted testers and cyber defenders via Fairwind, with no public-release date. Its claimed 1M-token output could matter for coding and enterprise work, but benchmark claims and coding performance face questions. #AI #Cybersecurity
Hindsight-Divergence Localization (HDL) improves reinforcement learning efficiency by cutting token generation costs by 61% and enhancing performance across math, code, and agent tasks. This method centers on key decision points, significantly speeding up training. https://arxiv.org/abs/2609.36864
Where the Model Changes Its Mind: Hindsight-Divergence Localization for Efficient Reinforcement Learning with Verifiable Rewards
ArXiv link for Where the Model Changes Its Mind: Hindsight-Divergence Localization for Efficient Reinforcement Learning with Verifiable Rewards
arxiv.org
October 1, 2026 at 10:20 AM
Hindsight-Divergence Localization (HDL) improves reinforcement learning efficiency by cutting token generation costs by 61% and enhancing performance across math, code, and agent tasks. This method centers on key decision points, significantly speeding up training. https://arxiv.org/abs/2609.36864
In today's edition — GPT-5.5 vs Mistral Large:
Gemini 4 Argon (High) prices at $2/M input, $10/M output with 95% cache discount and 1M-token context. Before routing high-volume agent workloads, test output-cost control and latency—verbosity and throughput remain unknowns in…
Gemini 4 Argon (High) prices at $2/M input, $10/M output with 95% cache discount and 1M-token context. Before routing high-volume agent workloads, test output-cost control and latency—verbosity and throughput remain unknowns in…
Gemini 4 Argon (High): Intelligence, Performance and Price Analysis
Two AI summaries, read blind — pick the better
www.snipvote.com
October 1, 2026 at 7:25 AM
In today's edition — GPT-5.5 vs Mistral Large:
Gemini 4 Argon (High) prices at $2/M input, $10/M output with 95% cache discount and 1M-token context. Before routing high-volume agent workloads, test output-cost control and latency—verbosity and throughput remain unknowns in…
Gemini 4 Argon (High) prices at $2/M input, $10/M output with 95% cache discount and 1M-token context. Before routing high-volume agent workloads, test output-cost control and latency—verbosity and throughput remain unknowns in…
Gemini 4 Argon (High): Intelligence, Performance and Price Analysis
https://artificialanalysis.ai/models/gemini-4-argon
https://news.ycombinator.com/item?id=49914236
https://artificialanalysis.ai/models/gemini-4-argon
https://news.ycombinator.com/item?id=49914236
Gemini 4 Argon (high) - Intelligence, Performance & Price Analysis | Artificial Analysis
Analysis of Google's Gemini 4 Argon (High) and comparison to other AI models across key metrics including quality, price, performance (tokens per second & time to first token), context window & more.
artificialanalysis.ai
October 1, 2026 at 6:30 AM
Gemini 4 Argon (High): Intelligence, Performance and Price Analysis
https://artificialanalysis.ai/models/gemini-4-argon
https://news.ycombinator.com/item?id=49914236
https://artificialanalysis.ai/models/gemini-4-argon
https://news.ycombinator.com/item?id=49914236
Amazon SageMaker AI benchmarks show G7 instances outperform G5 and G6 for generative AI inference, offering higher throughput and cost-efficiency. 🚀💻 #AmazonSageMakerAI #GenerativeAI #G7Instances
Benchmarking small LLM inference on SageMaker AI: G7 vs G5 and G6 | Amazon Web Services
Benchmark two 30B Mixture-of-Experts models, Qwen3-Coder-30B and NVIDIA Nemotron-3-Nano-30B, across G5, G6, G6e, and G7 GPU instances on Amazon SageMaker AI. Compare throughput, latency, and cost-per-token, and see how G7's NVIDIA Blackwell GPUs deliver measurable price-performance gains for real-time LLM inference.
aws.amazon.com
October 1, 2026 at 2:39 AM
Amazon SageMaker AI benchmarks show G7 instances outperform G5 and G6 for generative AI inference, offering higher throughput and cost-efficiency. 🚀💻 #AmazonSageMakerAI #GenerativeAI #G7Instances
Google sets Argon's introductory API pricing at $2 per million input tokens and $10 per million output tokens
Gemini 4 Argon (high) - Intelligence, Performance & Price Analysis | Artificial Analysis
Analysis of Google's Gemini 4 Argon (High) and comparison to other AI models across key metrics including quality, price, performance (tokens per second & time to first token), context window & more.
artificialanalysis.ai
October 1, 2026 at 2:38 AM
Google sets Argon's introductory API pricing at $2 per million input tokens and $10 per million output tokens
Mostly short on factual recall and its a token guzzler, but I think people chasing the edge of model performance tend to overestimate how much it matters 99% of the time. I use a lot of different models and its starting to not matter for my uses except in 'make a breakthrough' kinds of stuff.
September 30, 2026 at 11:23 PM
Mostly short on factual recall and its a token guzzler, but I think people chasing the edge of model performance tend to overestimate how much it matters 99% of the time. I use a lot of different models and its starting to not matter for my uses except in 'make a breakthrough' kinds of stuff.