Renat Aksitov

[Renat Aksitov @Korvin79]

@sergeykarayev You can try go with Hawking's “the universe would have neither beginning nor end; it would simply be.”

[Renat Aksitov @Korvin79]

@_arohan_ Hardly a drama, Arnold's principle in action :)

[shreya rajpal @ShreyaR]

the only things to look for when hiring AI/ML people is
- do they have a good sense of how to evaluate if something (model/system) works. can they go really deep into setting up evals, data, cohorts, etc.
- do they set up a high throughput experimentation harness
- do they keep up with research news

people think that good research is about getting good ideas, but ideas are really cheap and a good system for experimentation and actually, truly evaluating if your ideas are good or bad will already put you in the 90th percentile

[Taco Cohen @TacoCohen]

What I look for when hiring? EXTREME PARANOIA about code and data

[Omar Khattab @lateinteraction]

Since there seems to be another “DSPy finally clicks for me, how were they saying this 2+ years ago” moment right now,

You might wonder if timing is an accident.

Consider reading this blog post I wrote last year On Impactful AI Research, especially guidelines 2, 3, and 4.

[Xinyun Chen @xinyun_chen_]

Super glad to see a lot of excitement about our course! Again, huge thanks to @denny_zhou for coming to Berkeley and sharing insights on LLM reasoning!! Please join us on 2nd lecture, @ShunyuYao12 will give an overview of LLM agents and share his thoughts on important directions.

[Denny Zhou @denny_zhou]

My talk slides: https://t.co/SQ3mw8NW6B on key ideas and limitations of LLM reasoning, largely based on work in GDM. Thank you so much for the invitation, Dawn!

[Raymond Hettinger @raymondh]

#python tip:
>>> infinite_defaultdict = lambda: defaultdict(infinite_defaultdict)
>>> d = infinite_defaultdict()
>>> d['x']['y']['z'] = 10

[Renat Aksitov @Korvin79]

@gooby_esq @jobergum If they using RRF for hybrid, then not very surprising. If a linear combination, than simply did not tune a coefficient? (tuned should be better by definition)

[Andrew Ng @AndrewYNg]

The effort to protect innovation and open source continues. I believe we’re all better off if anyone can carry out basic AI research and share their innovations. Right now, I’m deeply concerned about California's proposed law SB-1047. It’s a long, complex bill with many parts that require safety assessments, shutdown capability for models, and so on.

There are many things wrong with this bill, but I’d like to focus here on just one: It defines an unreasonable “hazardous capability” designation that may make builders of large AI models liable if someone uses their models to do something that exceeds the bill’s definition of harm (such as causing $500 million in damage). That is practically impossible for any AI builder to ensure. If the bill is passed in its present form, it will stifle AI model builders, especially open source developers.

Some AI applications, for example in healthcare, are risky. But as I wrote previously, regulators should regulate applications rather than technology.
- Technology refers to tools that can be applied in many ways to solve various problems.
- Applications are specific implementations of technologies designed to meet particular customer needs.

For example, an electric motor is a technology. When we put it in a blender, an electric vehicle, dialysis machine, or guided bomb, it becomes an application. Imagine if we passed laws saying, if anyone uses a motor in a harmful way, the motor manufacturer is liable. Motor makers would either shut down or make motors so tiny as to be useless for most applications. If we pass such a law, sure, we might stop people from building guided bombs, but we’d also lose blenders, electric vehicles, and dialysis machines. In contrast, if we look at specific applications, like blenders, we can more rationally assess risks and figure out how to make sure they’re safe, and even ban classes of applications, like certain types of munitions.

Safety is a property of the application, not a property of the technology (or model), as @random_walker and @sayashk have pointed out. Whether a blender is a safe one can’t be determined by examining the electric motor. A similar argument holds for AI.

SB-1047 doesn’t account for this distinction. It ignores the reality that the number of beneficial uses of AI models is, like electric motors, vastly greater than the number of harmful ones. But, just as no one knows how to build a motor that can’t be used to cause harm, no one has figured out how to make sure an AI model can’t be adapted to harmful uses. In the case of open source models, there’s no known defense to fine-tuning to remove RLHF alignment. And jailbreaking work has shown that even closed-source, proprietary models that have been properly aligned can be attacked in ways that make them give harmful responses. Indeed, the sharp-witted @elder_plinius regularly tweets about jailbreaks for closed models. Kudos also to Anthropic’s @cem__anil and collaborators for publishing their work on many-shot jailbreaking, an attack that can get leading large language models to give inappropriate responses and is hard to defend against.

California has been home to a lot of innovation in AI. I’m worried that this anti-competitive, anti-innovation proposal has gotten so much traction in the legislature. Worse, other jurisdictions often follow California, and it would be awful if they were to do so in this instance.

SB-1047 passed in a key vote in the State Senate in May, but it still has additional steps before it becomes law. I hope you will speak out against it if you get a chance to do so.

[Original text (with links): https://t.co/MOQqFF6cID ]

[Xingyao Wang @xingyaow_]

Introducing OpenDevin CodeAct 1.0 - a new State-of-the-art open coding agent! It achieves a 21% unassisted resolve rate on SWE-Bench Lite, a 17% relative improvement above the previous SOTA by SWE-Agent.
Check out our blog or the thread 🧵for more details: https://t.co/qqAlqIqtj4

[Sharat @sharat_sc]

TIL: Langfun. This amazing library based on PyGlove from @dypeng et al supports symbolic and object oriented data programs using LLMs. Supports recursive type definition and calls that are very suitable to dynamic agentic workflows. https://t.co/MIEGGe7yUk

[Renat Aksitov @Korvin79]

@cutezu_ The problem asks whether solutions always exist and whether these solutions are smooth for all times.

The concern here is about the mathematical properties of the equations themselves, rather than solving them for specific conditions.

[Ahmad Beirami @abeirami]

A periodic reminder to reviewers:

If you ask authors for more experiments, then you need to communicate a clear hypothesis you're trying to verify with those (e.g., effectiveness on imbalanced data, generalization beyond a certain modality, scalability, etc).

Otherwise don't!

[Denny Zhou @denny_zhou]

For those who interpret/cite self-consistency (SC) as majority voting (MV) or other MV equivalents, they need to take an entry-level course in machine learning.

In our original paper (https://t.co/BA5MUzgBFp), Xuezhi and I have provided the math theory underlying SC: integrating out all possible reasoning paths to compute the full probability of the final answer. In terms of machine learning, the reasoning paths are latent variables, and this process is formally called marginalization. It is empirically done by sampling. That is all we need to know to understand and correctly apply SC. If one only looks at the first page of our paper, they would miss the math explanation part.

For speech recognition researchers/engineers, they would immediately see that SC shares the same spirit as CTC , one of the most beautiful machine learning methods that I have known. If there is no latent stuff, using SC is meaningless at least in theory, and the greedy outputs are optimal. This explains why it isn't often to see ideas similar to SC in the literature.

[Renat Aksitov @Korvin79]

@abeirami Difficulty of publishing negative results is part of the problem here

[Renat Aksitov @Korvin79]

@jobergum Any retrieval work that doesn't have BM25 baseline numbers is non-serious at best

[Xinyun Chen @xinyun_chen_]

Please consider submitting your work on LLM agents to our ICLR workshop (https://t.co/nO5gJbvyZm)! The submission deadline is Feb 3, 2024.

[AK @_akhaliq]

ReST meets ReAct: Self-Improvement for Multi-Step Reasoning LLM Agent

paper page: https://t.co/1hfFhbaCGo

Answering complex natural language questions often necessitates multi-step reasoning and integrating external information. Several systems have combined knowledge retrieval with a large language model (LLM) to answer such questions. These systems, however, suffer from various failure cases, and we cannot directly train them end-to-end to fix such failures, as interaction with external knowledge is non-differentiable. To address these deficiencies, we define a ReAct-style LLM agent with the ability to reason and act upon external knowledge. We further refine the agent through a ReST-like method that iteratively trains on previous trajectories, employing growing-batch reinforcement learning with AI feedback for continuous self-improvement and self-distillation. Starting from a prompted large model and after just two iterations of the algorithm, we can produce a fine-tuned small model that achieves comparable performance on challenging compositional question-answering benchmarks with two orders of magnitude fewer parameters.

[Renat Aksitov @Korvin79]

@lateinteraction @Moi39017963 @stalkermustang We do use Bamboogle auto-eval for selecting certain hyper-parameters (like temperature), which motivates introduction of the 2nd dataset as a control

[Renat Aksitov @Korvin79]

@lateinteraction @Moi39017963 @stalkermustang We don't use metric for constructing few-shot prompts. Just demonstrations to show correct / expected behavior for the specific step. Maybe, "vibes" are the metric in this case? :)