Ryan Allred (@ryanallred) header
Ryan Allred (@ryanallred) profile picture

Working to bring foundational AI innovations to market @subquadratic

Utah, USA
Joined April 2009
Gonna get GTA VI so that I can go bird watching..
8
Ryan Allred retweeted
This might be the craziest rocket video ever tbh
97
730
80
6,669
289,901
Ryan Allred retweeted
Update: this might be the worst NCAA Tournament bracket I’ve ever seen Who could have seen this coming…
Four of the best soccer teams in the west should not have to play each other in the first round of the NCAA Tournament Don’t be lazy @NCAA
16
8
2
149
36,723
Subquadratic’s SubQ Snap 1.2 gains 5 points over SubQ Snap 1.1 on the Artificial Analysis Decision Index v2.1, and is tied with Drex Max v1.1 and GLiDE v1.2 as the leading decision model SubQ Snap 1.2 is the latest decision model from @subquadratic. Pricing is unchanged at $0.04 per 1M input tokens, and it supports a 384K token context window. Decision Index v2.1 updates AA-Symmetry to v1.1. AA-Symmetry measures whether a model keeps the same decision and confidence when a question changes in ways that should not matter. v1.1 moves the answer options into the state and rewords them with the same meaning. Key takeaways: ➤ SubQ Snap 1.2 scores 73 on the Artificial Analysis Decision Index v2.1, up from 68 for SubQ Snap 1.1: Drex Max v1.1 (73) and GLiDE v1.2 (73) are within 0.1 points. The largest gains are on AA-Paperclip (+8 points) and AA-Symmetry (+8 points). ➤ The AA-Paperclip gain comes from more documents answered fully correctly: the all-pass rate rises from 46% to 62%, level with Drex Max v1.1 (63%) and GLiDE v1.2 (62%). ➤ The gains on AA-Omniscience-Decision and HLE-MultiChoice come from abstention, not accuracy: non-hallucination rates rise from 85% to 89% and from 46% to 55%. Accuracy is flat at 21% and 9%. On AA-Omniscience-Decision, SubQ Snap 1.2 attempts 29% of questions, down from 33%.
17
13
3
127
26,911
My “For You” timeline is absolute garbage right now. So sad.
2
2
283
One of the most common critiques of decision models is that they change their answer if you change the order of the options. But not anymore! SubQ Snap v1.2 now has “permutation invariance” built into the model so that it gives the same answer no matter the order.
Today we’re releasing SubQ Snap 1.2 - the smartest and most capable decision model for enterprise engineers. Tied for #1 on the @ArtificialAnlys Decision Index, Snap has leveled up in 4 core areas: Intelligence - Increased from 68 to 73 on the AA decision index. Order Sensitivity - Choice questions now return identical decisions regardless of option ordering nearly 100% of the time. Long-Context Performance - Increased from 63.2 to 68.4 in multi-document reasoning, outperforming Jev (50.5) and d1-ultra (exceeds 32k limit). Decision Accuracy - Snap scores a 69.1 on the open source component of the Hugging Face Decision index, outperforming Jev 1.14 (60.7), and d1-ultra (68.1). The 384k-token context window, multimodality, request format, and $0.04 per 1M tokens pricing remain identical to Snap 1.1. To start building with 100M free tokens, visit subq.ai
1
15
426
Today, we released SubQ Snap 1.2. It is the only decision model where the answer doesn’t change when the option order does. It is also the only model with a sub-quadratic model architecture, enabling a 384k-token context window without high latency. Our decision models are being built not for model judging or routing but doing actual LLM work, like tool calls, context curation, or unstructured document analysis. To start building with 100M free tokens, visit subq.ai. P.S. We have started to give some Snap users access to the latest version of our super long-context LLM.
Today we’re releasing SubQ Snap 1.2 - the smartest and most capable decision model for enterprise engineers. Tied for #1 on the @ArtificialAnlys Decision Index, Snap has leveled up in 4 core areas: Intelligence - Increased from 68 to 73 on the AA decision index. Order Sensitivity - Choice questions now return identical decisions regardless of option ordering nearly 100% of the time. Long-Context Performance - Increased from 63.2 to 68.4 in multi-document reasoning, outperforming Jev (50.5) and d1-ultra (exceeds 32k limit). Decision Accuracy - Snap scores a 69.1 on the open source component of the Hugging Face Decision index, outperforming Jev 1.14 (60.7), and d1-ultra (68.1). The 384k-token context window, multimodality, request format, and $0.04 per 1M tokens pricing remain identical to Snap 1.1. To start building with 100M free tokens, visit subq.ai
9
10
3
49
3,372
Will the hand thing show when I post this tweet?
1
2
276
Looks like I got randomly assigned to the other group in the A/B test
1
1
1
124
Ryan Allred retweeted
The bitter lesson has always been not just a push to leverage computation but a critique on methods. Methods that generalize and scale better win, and modeling those methods after human intuition is a waste of time. Somehow, much of the world forgot that methods matter. Many of our methods today do not generalize or scale well. The results: - tapping out the world’s resources extremely early in the deployment of AI - increasing intelligence requires scaling model size faster than compute efficiency can offset it - humans are the primary bottleneck in AI deployment - demand for human-curation of knowledge is so extreme that any company with access to a human workforce can quickly generate millions in data revenue - robotics intelligence is limited and task-specific The bitter lesson requires us to rethink our methods so that they better leverage computation and generalize, and it explicitly shows an evolution of algorithms over time. Believing that the bitter lesson expects us to indefinitely scale a single static method is incredibly unintellectual. Here are some methods that need attention: - training and inference without quadratic compute complexity - removing the KV cache - something better than massive pre-training with BCE only - is backpropagation still the best learning method? - in-context learning for robotics with massive memory efficiency improvements - algorithm-hardware co-development - how do we remove humans from harness development?
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
2
4
1
22
1,691
Ryan Allred retweeted
SubQ is sponsoring hackathon this weekend in SF. We will be giving early access of our unreleased models as well! 👀
San Francisco 🌁 We’re hosting a hackathon for the @aidotengineer Code Summit, in partnership with @GoogleDeepMind, @MetaforDevs, @MongoDB, @Fetch_ai and @subquadratic. 🏆 $11,000 in cash prizes 
📅 Nov 7 at Shack15 Apply below 👇
1
7
128
Ryan Allred retweeted
Introducing SubQ Snap 1.1 on Baseten. Built for long-context decision-making, SubQ Snap 1.1 by @subquadratic has a 384K context window. It scores 65 on the Artificial Analysis Decision Index, tied for #1. Explore SubQ Snap 1.1 on Baseten: baseten.co/library/subq-snap…
4
7
2
63
6,005
Ryan Allred retweeted
SubQ Snap is now available on @baseten.
2
11
1
31
5,209
Ryan Allred retweeted
Demand for our long context decision model is stronger than anticipated, so we’ve partnered with @baseten to expand access. Big thanks to @amiruci and the team. Look out for another fun update this week. teapawt.net/subquadratic/status/21…
SubQ Snap is now available on @baseten.
2
6
28
2,108
Ryan Allred retweeted
I think that AA should remove Max mode results. It’s not how these models are actually used and it inflates scores and tokens per task which leads to a lot of misinformation.
Announcing Artificial Analysis Coding Agent Index v2.0: FrontierCode v1.1 from Cognition replaces DeepSWE, SWE-Atlas-QnA grading is revised, and Claude Opus 5.5 in Claude Code leads at 61.7 Today we are launching v2.0 of the Coding Agent Index. The Index evaluates models together with their agent harnesses, now across 290 tasks covering software engineering, terminal use, and repository Q&A, with the three benchmarks weighted equally. Changelog (Index v1.5 → Index v2.0): → FrontierCode v1.1 replaces DeepSWE v1.1. FrontierCode is a benchmark from @cognition with a private test set: 100 software engineering tasks on real open-source repositories, graded on whether a maintainer would merge the change. An attempt that fails a blocking check scores zero, and one that passes them all scores between 70% and 100% depending on the non-blocking checks it meets. We also scan each trajectory with a modified version of Cognition's rule-based scanner, and any attempt flagged for accessing solution material scores zero. → Revised SWE-Atlas-QnA grading, now SWE-Atlas-QnA-AA. Scale AI's 124 repository Q&A tasks are unchanged. We reworded 127 of the 1,306 rubric criteria for more consistent verdicts and moved to GPT-6 Luna (xhigh) as the judge, with five votes per criterion. The dataset and judge configuration are published on Hugging Face. Key results: ➤ Claude Opus 5.5 leads the Index: Claude Code with Claude Opus 5.5 (max) scores 61.7, followed by Claude Code with Claude Sonnet 5.5 (max) at 59.3, Muse Code with Muse Wonder Full (max) at 58.9 and Claude Code with Claude Fable 5.1 (max with fallback) at 57.6. ➤ FrontierCode v1.1 is where Opus 5.5 pulls ahead: Opus 5.5 scores 56.0% on FrontierCode, ahead of Codex with DeepSeek V4.1 Flash (max) at 52.6%, Codex with GPT-6.1 Sol (max) at 49.4% and Codex with GPT-6 Astra (max) at 49.3%. Below the top, FrontierCode scores are tightly packed: 15 of the 24 configurations on the Index score between 43% and 49%. ➤ SWE-Atlas-QnA-AA grading is stricter and more consistent: across the 21 configurations evaluated on both Index versions, the mean score falls 1.7 points, from 58.8% to 57.1%, with 14 of 21 moving down. Individual configurations move by up to 7 points: Muse Code with Muse Spark Open (xhigh) falls 7.0, Kimi K3 in Kimi Code CLI and Codex with GPT-6 Luna (max) 6.2, and Muse Code with Muse Wonder (max) 5.6, while Codex with GPT-6.1 Sol (max) gains 2.1. ➤ Four configurations sit on the cost frontier: Codex with DeepSeek V4.1 Flash (max) scores 47.3 at $0.39 per task, Codex with GPT-6.1 Sol (max) 54.1 at $1.50, Codex with GPT-6 Astra (max) 54.7 at $7.04 and Claude Code with Claude Opus 5.5 (max) 61.7 at $11.42. Costs are pay-per-token API costs, not subscription prices. We will keep updating the Coding Agent Index incrementally, with more evaluations and agents to come.
63
12
3
920
32,590
Give Saul a follow, one of the smartest guys in AI and he's new to X!
It's been exciting to work on this model alongside such a talented team. These models behave quite differently from traditional LLMs, which has pushed us to think creatively about how we optimize them. Really proud of the team. Check out SubQ Snap 1.1!
2
8
434
Subquadratic has released SubQ Snap 1.1, tying d1-ultra and Elephant 9B for first on the Artificial Analysis Decision Index v1.0, driven by a low hallucination rate SubQ Snap 1.1 is the new decision model from @subquadratic. It is priced at $0.04 per 1M input tokens and supports a 384K token context window, compared to 32K for d1-ultra and 16K for Elephant 9B. Key takeaways: ➤ SubQ Snap 1.1 scores 65 on the Artificial Analysis Decision Index v1.0, tied with d1-ultra and Elephant 9B: the three models are 8 points ahead of the next model, Jev 1.14 (57). ➤ SubQ Snap 1.1's AA-Omniscience-Decision result is driven by a low hallucination rate: its non-hallucination rate of 86% is second only to Elephant 9B (93%), ahead of d1-ultra (85%). Its accuracy is 20%, compared to 32% for d1-ultra and 12% for Elephant 9B. The non-hallucination rate is the share of questions not answered correctly where the model chose "I don't know". ➤ SubQ Snap 1.1 scores 18% on HLE-MultiChoice: level with Elephant 9B (18%) and ahead of d1-ultra (15%). HLE-MultiChoice is the text-only multiple-choice subset of Humanity's Last Exam. ➤ SubQ Snap 1.1 costs $30 per 1M tasks: comparable to d1-ultra ($32), and ~3.6x the cost of Elephant 9B ($8.4), which scores the same 65.
15
10
7
165
21,233
Snap’s long-horizon agentic tool use capability comes in handy when playing Super Smash Bros!
3
2
2
14
594
I'm stoked for the public to now have access to a SubQ model that's top of its class and that uses our innovative Subquadratic Sparse Attention (SSA)!
Today we're introducing the smartest, most capable version of our enterprise-ready decision model: SubQ Snap 1.1. Snap is tied for #1 on the Artificial Analysis Decision Index and has made big gains on agentic tasks, outperforming Jev by +7.2 points across 16 leading enterprise benchmarks. All while maintaining 384k tokens of context, in a single pass, as the only model with constant compute per token. - Powered by Subquadratic Sparse Attention. - Tied for #1 on the @ArtificialAnlys Decision Index. - 96.2 on When2Call, up from 75.2, and up 6 points on @huggingface 5 Tools & Automation benchmarks. It knows when to call a tool, when to ask, and when to decline. - 81% fewer confident errors on questions it can't answer. - $0.04 per 1M input tokens. No charge for output. Go to subq.ai for access, and 100M free tokens.
1
25
749