Scaled Cognition
@ScaledCognition
Followers
564
Following
62
Media
4
Statuses
54
The only frontier model for CX that eliminates hallucinations.
SF, NYC, Boston
Joined January 2023
🚀 New Research from Scaled Cognition TL;DR: Training speedups of up to 70x on tree-structured data. Not 70%. _70x_. We just published a blog on Prompt Trees: Training-time Prefix Caching, a technique that delivers up to 70× training speedups on tree-structured data. By
scaledcognition.com
We present a method for encoding tree-structured data (like you might get from conversation rollouts while doing RL training for conversational agents) using a standard transformer. This method gives...
0
1
6
THE MODEL CONSTELLATION GAMBIT - Because generalist models are non-deterministic, AI application layer companies cannot trust the output. To compensate, they build Constellations, complex model chains where a router classifies the input, a frontier model creates a draft, a
0
2
7
Increasingly, AI application-layer companies present complex model chains as a flex—proof of sophistication, and a reason enterprises should believe they couldn’t possibly build this themselves. In reality, these Rube Goldberg–style Constellations are an admission that the
scaledcognition.com
The Constellation approach promises reliability by chaining AI models—but in practice it multiplies errors, latency, cost, and operational risk.
0
0
1
If people are already building AI agents on top of frontier LLMs, why do we need a new architecture at all? That’s the question our CTO, @profdanklein, was asked during a recent talk at the @FinRegLab AI Symposium. LLMs are powerful and flexible. But their most visible
0
0
2
Attending @NeurIPSConf last week with some of the @ScaledCognition research team was a blast. Met lots of cool new people and had many fun conversations. "What if models didn't hallucinate" was what drew folks to our booth (the hint is that adding structure helps but happy to
2
4
15
Because if 74 percent of agents require human babysitting, scale is not just expensive. It is impossible. The future is not agents that take ten steps instead of five. The future is systems that can think, verify, iterate, and improve before a human ever needs to intervene.
arxiv.org
AI agents are actively running in production across diverse industries, yet little is publicly known about which technical approaches enable successful real-world deployments. We present the first...
0
0
0
Scaled Cognition was designed for this gap from day one. We treat autonomy not as a gimmick but as a capability layer — one that provides reliability, safety, correctness, and makes intelligent systems actually work in the real world.
1
0
0
In other words: Most agents today are not autonomous. They are constrained systems with hard ceilings. The gap between what organizations require and what current agents deliver is now quantifiable and impossible to ignore.
1
0
0
306 practitioners. 20 in-depth case studies. 26 industries. The data tells a very clear story: 🔹 68% of agents can execute 10 steps or fewer before a human must intervene 🔹 70% rely on prompting off-the-shelf models rather than trained systems 🔹 74% depend on human evaluation
1
0
0
Most people think AI agents are already autonomous. This new research shows they are not. Not even close. A first-of-its-kind study out of Berkeley and Stanford just benchmarked how agents are used in the real world.
1
2
7
Always a pleasure partnering with teams who move this fast. Appreciate you, @basetenco 🚀
Agents that don't hallucinate? Meet APT: @ScaledCognition's Agentic Pretrained Transformer — the only frontier model for CX that eliminates hallucinations. We've been partners (and fans) of the Scaled Cognition team from launch day to massive scale, working with their engineers
1
1
4
I read somewhere that parenting is really just prompt engineering. As parents to two teenagers we’re constantly trying to figure out which token sequence will actually work to elicit the desired behavior, and which sequences will stick for more than ten minutes to get the model
0
4
8
Excited to be at @NeurIPSConf this week, and to share some of the cool stuff we've been building for reliable agents at @ScaledCognition at our booth in the expo hall (demo included!). We're also hiring :)
1
2
5
This is the path to hallucination-free CX and it’s why leading BPOs and brands trust APT-1, our frontier model that eliminates hallucinations, to run their CX.
1
0
1
The next frontier isn’t bigger models. It’s stable models. Purpose-built. Domain-grounded. Able to follow instructions with precision.
1
0
1
General-purpose AI sounds fluent but can’t be trusted when accuracy, compliance, or money movement is on the line. You can’t build real workflows on Jell-O.
1
0
1
As @karpathy pointed out on @dwarkeshpodcast, today’s models spend their energy on “memory work instead of cognitive work” because the internet is such noisy data. Scaling slop still gives you slop.
1
0
1
Most AI looks solid until you put real weight on it. Lean on it in production and the whole foundation starts to wobble. Our CEO, @roth_dan breaks down why models trained on the chaos of the internet fail the moment the stakes are real and why the future belongs to specialized,
1
2
7
I'm headed to NeurIPS now, excited to see people! Also, ICYMI, we just published a way to represent RL rollouts as trees and encode them with a single pass of a standard transformer, giving 70x speedups to post-training:
New blog post - Prompt Trees: Training-time Prefix Caching. By the research team at @scaledcognition. TL;DR: Training speedups of up to 70x on tree-structured data. Not 70%. _70x_. https://t.co/EYD96dHAHk (preprint version coming soon)
0
1
11