Morph
@morph_labs
Followers
4,203
Following
1
Media
3
Statuses
20
Explore trending content on Musk Viewer
Eminem
• 250433 Tweets
River
• 209846 Tweets
Celtics
• 160772 Tweets
LeBron
• 153750 Tweets
Lakers
• 133492 Tweets
Fernando Valenzuela
• 110551 Tweets
#PetrichorTheSeries
• 90905 Tweets
Deyverson
• 84161 Tweets
Knicks
• 76828 Tweets
Gallardo
• 54350 Tweets
#iQIYIOriginal
• 53587 Tweets
政党助成金2000万円
• 44496 Tweets
非公認議員
• 39913 Tweets
Fonseca
• 31928 Tweets
प्रियंका गांधी
• 23012 Tweets
非公認候補の党支部
• 21050 Tweets
Daisuke
• 20590 Tweets
#Wayanadinte_Priyanka
• 20195 Tweets
支部の活動費
• 19432 Tweets
Anthony Davis
• 18914 Tweets
東京メトロ
• 17954 Tweets
DONNY 4everAKapamilya
• 16982 Tweets
Randle
• 15571 Tweets
Gobert
• 14938 Tweets
Tigres
• 11524 Tweets
PRABOWO-GIBRAN KerjaFOKUS
• 11463 Tweets
SiapBEKERJA UntukRAKYAT
• 11209 Tweets
気圧のせい
• 11204 Tweets
#わたしの推し牛
• 10371 Tweets
We have developed a method called self-teaching for robustly adapting chat LLMs to new knowledge bases. Self-taught models forget less, remember more, and are better at multistep reasoning over learned facts.
3
43
247
We have raised 5.75M from
@khoslaventures
and others to build infrastructure for the next billion artificially intelligent software engineers.
We solve some of the hardest problems in the world around developing and deploying AI SWEs, and we're hiring.
11
15
125
On a challenging multi-hop Q&A benchmark released on
@huggingface
, we observe that in comparison to RAG / long-context / finetuning, self-taught models have better closed-book multi-document reasoning, less forgetting, and better in-context reasoning.
1
7
35
Systematically evaluating AI coding assistants is difficult. Morph is developing infrastructure for automated codebase-specific evaluations, real-time semantic code search, and on-the-fly synthetic data used by companies like
@togethercompute
1
2
23
We're thrilled to be working with other great investors like
@betaworks
,
@kokoxsu
, Irregular Expressions, Roar Ventures and incredible angels like
@prasanna
,
@amasad
,
@vipulved
,
@ccrisccris
,
@ChrSzegedy
1
0
14
At Morph, we believe that effective infrastructure for AI SWE agents must be developed with a fundamental empathy for the machine. We are a small, interdisciplinary, and extremely fast-moving team:
0
0
9
Our experiments focus on the well-studied but relatively weak (by current standards) Llama 2 Chat 7B, but we have observed that performance only improves when using stronger chat language models to self-teach.
1
0
8
Self-teaching is robust, effective, and presents an Pareto improvement upon finetuning for bootstrapping knowledge into a model for downstream question-answering, even when the questions are off-distribution or require reasoning across multiple documents from memory.
1
0
8
AI SWE agents will soon vastly outnumber human software engineers. We will lead the way in the development of AI-centric infrastructure to support this new category of synthetic knowledge workers and the companies that develop and operate them.
1
0
7
Codebase Q&A systems optimized for our automated evaluations record SOTA human-rated helpfulness scores in a blind trial on a handwritten SWE-bench-style benchmark derived from resolved GitHub issues over 5 prominent open-source codebases
1
0
5
Customer highlight: congrats
@StoriaAI
on this dataset release! This data was generated with our automated codebase-specific eval platform at 25X lower cost and 10X faster than other leading data providers.
Tired of building AI coding agents based on vibes?
Together with our friends from
@morph_labs
we made a real-world dataset that gives us a ladder to climb: 1,000 questions about the Transformers library.
Here are our initial learnings about proprietary APIs. 🧵
1
1
13
1
0
4
With human judgment as our north star, Morph Labs is working to make useful, cheap, scalable and codebase-specific benchmarks possible
1
0
3
Morph and
@togethercompute
are collaborating on research aimed at enhancing end-to-end RAG system performance for codebases through leveraging automated evaluations, superior search and high-quality synthetic data. Read more in their blog here:
2
0
3
Message us at hello
@morph
.so to work with us on deploying smarter and more robust chat assistant LLMs for your domain-specific use-case.
0
0
2
Automated evals like ours enable faster development of AI coding systems while addressing the high cost of manually curating benchmarks of questions over diverse codebases:
0
0
2
Our automated evaluations identify higher-performing codebase Q&A systems and cause optimized systems to attain higher average human-rated helpfulness. Because they’re codebase-specific, our benchmarks can measure design choice effect in end-to-end RAG systems for any repository
0
0
1