Solomonic learning: Large language models and the art of induction

Large language models’ emergent abilities are improving with scale; as scale grows, where are LLMs heading? Insights from Ray Solomonoff’s theory of induction and stochastic realization theory may help us envision — and guide — the limits of scaling.

“One year of research in neural networks is sufficient to believe in God.” The writing on the wall of John Hopfield’s lab at Caltech made no sense to me in 1992. Three decades later, and after years of building large language models, I see its sense if one replaces sufficiency with necessity: understanding neural networks as we teach them today requires believing in an immanent entity.

Stefano Soatto.png
Stefano Soatto, a vice president and distinguished scientist with Amazon Web Services.
Credit: UCLA Samueli

Let’s start from the basics: when we teach machine learning, we say that memorization is bad, because it leads to overfitting and prevents generalization. Generalization is good — so good that, to achieve it, we incentivize machines not to memorize, through “regularization”. We even prove theorems — so-called uniform generalization bounds — that guarantee generalization no matter what distribution the data are drawn from, provided we avoid memorization.

But my mother always told me not to generalize, and she had me commit to memory countless useless poems in elementary school. Why am I teaching that generalization is good and memorization is bad, when I was taught the opposite?

Biology vs. technology

Machine learning has historically drawn inspiration from biology. But biological systems have hard ontogenic and phylogenic memory bounds: our synapses cannot memorize everything we experience, and our DNA cannot transmit the knowledge we’ve accumulated to our descendants. (As an educator and father, I often wished I could upload what I have learned into my students and kids. I haven’t figured that one out, but can we at least do it for AI models?) Furthermore, biology imposes a strong evolutionary bias toward minimizing inference latency: when facing an animal in the wild and having to determine who’s whose meal, we can’t reason through all past memories lest the decision be made for us.

In other words, biological systems are forced to adopt inductive learning, using specific data from the past (or a “training set”) to devise a process for handling any future data. Success in inference from inductive learning (or more simply, induction) relies on the so-called inductive hypothesis, that past performance can guarantee future rewards (the primate species called “financial advisor” has evolved out of this belief).

Related content
New method leverages vision-language models to formalize a comparison that had previously required human judgment.

Technology does not have the limitations of biological systems: there are no hard memory bounds (we can always add more storage) and no hard computational bounds (we can fire up more computers), at least until we hit cosmic limits. If we accept that machines do not have the same limitations as biology, what is the best inference paradigm for them? That is, given a training set and a test query, how can they devise the best answer?[1] If we want our model to operate in the constantly evolving real world, we shouldn’t assume the existence of a single distribution from which all data are drawn, in principio, nunc, et semper.

Inference that allows processing the training data at inference time is called transductive inference, or transduction. Transduction calls for us to memorize and reason, unlike induction, which wants us to generalize and forget. To perform optimal inference with respect to any hypothetical distribution in the future, one must memorize past data and, only when presented with a specific query, deploy “reasoning” skills and access memory to compute the best possible answer to that query.

Induction calls for forgetting what does not matter during training, under the assumption that the training set is representative of all future data. But in reality, one cannot know what data will be useful when, so memorization is wise if one can afford it, even when the data — like the writing on John Hopfield’s lab’s wall — does not make sense in that moment.

Transductive inference from inductive learning

Uniform generalization bounds may seem powerful because they are valid for any distribution; but for them to work, there can be only one distribution from which both past and future data are independently sampled. Paraphrasing the statistician Bruno de Finetti, this distribution does not exist in any objective or material sense. It is an abstract concept, the product of our imagination. Something we concoct to guide our intuition and analysis.

Related content
In addition to its practical implications, recent work on “meaning representations” could shed light on some old philosophical questions.

The inductive hypothesis is fundamentally not verifiable: any finite training data could have been drawn with identical likelihood from infinitely many distributions, so even if there was a single true one, how would we know which? Once the present is past, we cannot repeat the experiment. The inductive hypothesis is a statement of faith and uniform generalization bounds an expression of hope, not quite within the scientific realm.

Don’t get me wrong: hope can pay off. The future often does resemble the past. But many of the mechanisms that generate the data we care about today, in business, finance, climate, and language, evolve over time. The same word can carry a different meaning today than it did a century, or even a decade, ago. The point is that whether the inductive hypothesis holds or not cannot be known ahead of time.

Solomonoff inference

What if we forgo generalization and embrace memorization and reasoning? Is that what LLMs are doing? If so, where are they heading? What does the limit of optimal transductive inference look like?

The answer was given in 1964 by the mathematician Ray Solomonoff and is now known, somewhat confusingly, as Solomonoff induction. I will refer to it as Solomonoff inference, which can be thought of as the limit of scaling laws when we allow memory, computational capacity, and time to grow to infinity.

Solomonoff inference is optimal with respect to all computable distributions, averaged with respect to the universal prior. The Church-Turing thesis predicates that any physically realizable mechanism belongs to this class. While infeasible in practice, since it requires infinite resources, Solomonoff’s algorithm is quite simple: execute all programs in increasing order of length until one manages to spit out all the data observed up to now, bit by bit, if it terminates.

Related content
The surprising dynamics related to learning that are common to artificial and biological systems.

The optimal algorithm is basically a lookup table with a switch. There is no insight, no knowledge, not even learning. If presented with the same query twice in a row, the optimal algorithm would repeat the same procedure all over, having learned nothing from past experience.

Solomonoff inference is quite unlike neural networks, which are trained by comparing gradient vectors in a high-dimensional space, where the data are embedded. But could it be that, as we scale LLMs to larger and larger sizes, their behavior is beginning to resemble Solomonoff inference? After all, LLMs are known to memorize, albeit imperfectly, and they can perform universal computation, at least if augmented with a scratchpad. Indeed, LLMs are already able to perform rudimentary transductive inference, now known as “in-context learning” — somewhat confusingly, as it involves no learning: if presented with the same context twice, an LLM would repeat the same process, with no improvement from experience.

So, if LLMs were to begin to perform Solomonoff inference, would they become “superintelligent”? Given no accepted definition of intelligence, let alone its superlatives, many tacitly assume inference performance as its proxy: “smarter” models (or students) perform better on tests, whether the SAT, GRE, or BAR, or the famed IMO math competition. The higher the score, the more “intelligent” the model must be! But the absolute best would be Solomonoff’s algorithm, and no matter what one’s definition of intelligence is, Solomonoff’s algorithm cannot meet it: if by mistake the IMO printed each question twice, Solomonoff’s algorithm would redo the same work twice, not exactly what most would call “intelligent” behavior.

As an analogy, an “inductive student” is a diligent pupil who studies the textbook and completes all homework assignments and practice problems before showing up at the exam. So long as the questions are close enough to practice problems, the inductive student does well. On the occasional odd (or out-of-distribution, as a believer in induction would say) question, the inductive student may not do as well.

By contrast, the “transductive student” does not study at all and instead shows up at the exam with the textbook in hand. Only after reading the first question does the transductive student go through the book to find all the pieces needed to assemble an answer. The student could, in principle, repeat the exercise all the way to the last question, learning nothing in the process. As Solomonoff showed us, there is no need to be smart if one has unbounded time, memory, and computational power.

Do we want models that perform well on benchmark exams, or is the kind of “intelligence” we want something else? Fortunately, inductive and transductive inference are not mutually exclusive. In fact, their difference is quite subtle, as one could frame either as a special case of the other, and the two coincide when the data are independently and identically distributed.

Related content
Technique that mixes public and private training data can meet differential-privacy criteria while cutting error increase by 60%-70%.

What matters is that LLMs are inductively trained transductive-inference engines and can therefore support both forms of inference.[2] They are capable of performing inference by inductive learning, like any trained classifier, akin to Daniel Kahneman’s “system 1” behavior — the fast thinking of his book title Thinking Fast and Slow. But LLMs are also capable of rudimentary forms of transduction, such as in-context-learning and chain of thought, which we may call system 2 — slow-thinking — behavior. The more sophisticated among us have even taught LLMs to do deduction — the ultimate test for their emergent abilities.

AI models’ inferential abilities are improving organically with scale — although they’re still inferior to those of the best humans on most tasks. But they are also being actively fostered through the use of formal-verification tools such as LEAN, as is happening at AWS. One could call this paradigm Solomonic learning: embrace memorization and foster reasoning, yet do not eschew induction. Simple tasks that might benefit from past experience can be solved inductively, saving time and energy, but doing so requires “understanding” and “insight”.

Given that paradigm, the question is what classes of models best support Solomonic learning.

Architectures for Solomonic learning

Solomonic learning requires models that can memorize and perform computation at inference time, in addition to performing ordinary induction. The model architectures therefore need eidetic (verbatim) working memory, which could fade over time, to support computation; but they also need long-term memory to easily retrieve facts from the distant past (the purpose for which humans invented the printing press).

To adapt to changing conditions, they need their long-term memory to decay in synchrony with changes to the mechanisms that generate the data they process. Evolution does that for biological agents, to the benefit of the species rather than any one individual. Transformers, the workhorses of current LLMs, have eidetic (verbatim) memory “in context”, but only until tokens slide out of context. They also have permanent memory “in weights”, but training data are not accessible eidetically from the weights, and there is no long-term adaptation. Eidetic long-term memory can be accessed through RAG (retrieval-augmented generation), but in current Transformers, RAG is not integrated into the primary (autoregressive) inference loop.

Stochastic realization theory and input-dependent state space models

Half a century ago, stochastic realization theory tackled the question of how to model sequential data for downstream decision or control tasks. The “state” of the model was defined as the function of past data that is sufficient for the future, meaning that, given the state, one can discard all past data and predict future data as well as if the data had been retained.

The trivial state is the data itself. An optimal state, by definition, supports an optimal predictor, which is one that makes the prediction error unpredictable. Then, by construction, the state contains all the “information” in past data. During training, the states of LLMs are their weights, so it should be no surprise that next-token prediction is the method of choice for training them. During inference, the state of a Transformer-based LLM is the sliding window of tokens, which is “deadbeat”, meaning that it decays to zero in finite steps without a driving input.

B'MOJO.jpg
In B’MOJO, a state-space model (SSM) computes a fading memory that represents long-range dependencies through a fixed-dimensional representation (pink). The eidetic memory, by contrast, selects tokens from the past (dark-blue x's) using an innovation test over the SSM output and appends them to the current sliding window. Adapted from "B'MOJO: Hybrid state space realizations of foundation models with eidetic and fading memory".

In general, as we observe more and more data during both training and inference, the state must grow apace. In the 1970s, an unbounded state was unthinkable, so the key question was how to find a fixed-dimensional state that is optimal even as the data volume grows to infinity. Therefore, stochastic realization theory focused on Markov processes that admit a finite-dimensional state.

Since any finite-memory sequence could be modeled as the output of a linear model driven by white zero-mean Gaussian noise, the attention was all on linear state-space models (SSMs). While simplistic, such SSMs were good enough to take us to the moon. Today, an unbounded state is not unthinkable. Nonetheless, LLM weights are fixed after training, and the context size is imposed by hardware limitations. So we need richer architecture families.

As an aside, I wish to stress the distinction between the model, which is any state-space realization that supports optimal prediction (there are generally infinitely many), and the system, which is the “real” mechanism that generates the data. The system is unknown and unknowable; the model is tangible and entirely under our control. Although as engineers we are trained to believe that models of the world converge to the “true” system as they improve, this position — known in epistemology as "naïve realism" — is scientifically indefensible.[3]

Amazon’s Stefano Soatto on how learning representations came to dominate machine learning.

To stress the dichotomy between the system and the model, in 1979, Anders Lindqvist and Giorgio Picci derived an equation that, four decades later, is at the heart of diffusion models. In a dissipative physical system, time cannot be reversed, bu it can in a model of that system, for instance a Gaussian SSM. The structure of the reverse diffusion in the model is the same as the forward diffusion, a fact that is exploited in diffusion models for image generation.[4]

Unlike deadbeat Transformers, SSMs have unbounded memory, but it fades, making them incompatible with optimal transductive inference. Again in the 1970s, the late Roger Brockett triggered a burst of interest in input-dependent state-space models, where some of the parameters are affected by the input, the simplest case being when they interact (bi-)linearly with the state. Art Krener showed that such bilinear SSMs can approximate an arbitrarily complex nonlinear (smooth) model. Alberto Isidori and coworkers extended stochastic realization theory to bilinear models, but still with an eye to making the state as small as possible.

Even 30 years later, prior to the deep-learning revolution, when we used input-dependent SSMs to generate videos of dynamic textures, we were still focused on keeping the state dimension as small as possible, encouraged by the fact that 20 states were sufficient to animate and control the rendering of waterfalls, flames, smoke, foliage, talking faces, and other stationary processes. Thanks to the reversibility of the model, we could even make smoke or steam move faster, slower, or backwards!

Deep learning twisted Occam’s razor by trying to make the embedding dimension of the training state (the weights) as large as possible, not as small as possible. Dimension is only an upper bound on “information,” and the key to induction is to limit the “information” in, not the dimension of, the trained weights.[5] Two decades later, we stacked SSMs into a neural architecture by feeding the (input-dependent) prediction residual of one layer to the next.

A breakthrough came with Mamba, which showed that efficient implementation at the hardware level is key. When Mamba is stripped down (as it is in appendix E of our recent paper on architectures to support transductive inference), it is a stack of bilinear SSMs (which Mamba’s developers call “selective state-space models”) restricted to non-interacting states (diagonal dynamics), so it can be implemented efficiently in hardware.

Diagonal SSMs are disjoint from and complementary to Transformers. Autoregressive (AR) Transformers have nilpotent dynamics, meaning that the state transition matrix becomes zero in a finite number of steps in the absence of external input. Mamba has diagonal dynamics, and nilpotent matrices cannot be diagonalized. Diagonal SSMs support infinite fading memory; AR Transformers support finite eidetic memory, and neither is general. Instead, any general (bi-)linear system can be converted to a so-called canonical form, also derived in the 1970s, which can support both eidetic and fading memory.

Meet B’MOJO

B’MOJO is a family of architectures based on canonical realizations that include Transformers, Mamba-like SSMs, and any hybrid combination of the two. There are combinatorially many options, and the name of the game is to find those that are sufficiently general to support different memory regimes yet can be efficiently mapped to specific hardware in order to scale. We plan to release basic versions of B’MOJO both for GPU hardware and for Amazon’s Trainium hardware, so they can be easily compared with existing Transformers, SSMs, and hybrid architectures.

The writing on the wall

While a representation of the “true” system is fundamentally elusive, lending credence to the writing on the wall of John Hopfield’s lab back in 1992, building model realizations is a concrete exercise grounded in data. LLMs, where the “L” refers not to natural language but to the inner language that emerges in the trained model at scale, are stochastic realizations trained inductively as optimal predictors and coopted for (suboptimal) transductive inference and generation. If the training data subtend latent logical structures, as do sensory data such as visual or acoustic data, models trained as optimal predictors are forced to capture their statistical structure.

Related content
From the urgent challenge of "machine unlearning" to overcoming the problem of critical learning periods in deep neural networks, Alessandro Achille is tackling fundamental issues on behalf of Amazon customers.

Thus, LLMs in our parlance include so-called world models trained with visual, acoustic, olfactory, tactile, and other sensory data. The model is indifferent to whether tokenized data express some abstract concept in natural language or a physical measurement process in finite precision. The resulting LLMs can represent concepts and meanings, including physical concepts such as the laws of physics, and can in principle reason, although at present they appear to be mostly building ever bigger lookup tables. Regardless, as stochastic dynamical models, LLMs can be controlled, probed with causal interventions, made observable, and studied with the tools of dynamical-systems theory.

A model is an abstraction of the underlying world — not a representation of it, because there is no objective “it” to re-present, but a realization of it, made real through the only objective entity, which is the data. Synthetic data are just as real to the model as data produced by a physical measurement process, and aligning the two is the essence of perception, for this reason often referred to as controlled hallucination.

While much of the popular discourse denigrates hallucinations[6] as something to be avoided, the ability to hallucinate is necessary for reasoning. The question is not how to avoid hallucinations but how to control them, which is the process of alignment. Architectures designed for decision and control can help, and decades of work in dynamical systems and controls may provide insights — hopefully without the need to resort to divinity, as the writing on the wall suggested.

Footnotes

[1] Note that "best" does not mean "correct." If the data is insufficient to identify the correct conclusion, even the best answer can be wrong.

[2] The simplest form of inductive learning for transductive inference is transductive fine-tuning, a form of meta-learning: past data is used to "meta-train" a model that, at inference time, is fine-tuned with a small number of examples ("few shots") to perform a new task. LLMs take this program steps further, by using sequential data with a latent logical structure (not only natural language but also video, audio, and other signals) to produce an “inner language” (we call it "Neuralese") that can then be co-opted for transductive inference.

[3] Quoting Bertrand Russell: “We all start from 'naïve realism,' i.e., the doctrine that things are what they seem. ... The observer, when he seems to himself to be observing a stone, is really, if physics is to be believed, observing the effects of the stone upon himself. Thus science seems to be at war with itself: when it most means to be objective, it finds itself plunged into subjectivity against its will. Naïve realism leads to physics, and physics, if true, shows that naïve realism is false. Therefore naïve realism, if true, is false; therefore it is false.” Even the International Vocabulary of Metrology has dispensed with the notion of “true value” in its most recent revisions.

[4] In the paper that introduced diffusion models for image generation, the reverse-diffusion equation was attributed to a 1949 work of Feller. However, forward diffusion in the form in use today was not derived until 1960, so neither was reverse diffusion. Later references attribute the reverse-diffusion equation to a 1982 paper by B. D. O. Anderson, which, however, did not introduce it but instead described it, based on the 1979 paper of Lindqvist and Picci, correctly referenced in Anderson’s work, and extended it to more general models different from those in use in diffusion models today. The correct reference for the reverse-diffusion equation used in diffusion models is therefore Lindqvist-Picci 1979.

[5] I use quotes because defining information for the weights of a trained model entails some subtleties, but it can be done.

[6] "Hallucinations" are data generated by a model that are statistically compatible with the training set (in the sense of high likelihood under the trained model), yet "wrong", i.e., individually inconsistent with constraints that some external oracle has deemed "true" ("facts", or "axioms"). In other words, hallucinations are the product of any generative model. Outside formalized domains such as math or code, there is no objective "truth", so the oracle is replaced by an accepted knowledge base, which depends on the application. For "common sense" knowledge, the base is generally a large corpus of (more or less) verified facts, such as WikiData. Outside formalized domains, including the law, there is no guarantee that the facts or "axioms" are mutually compatible.

Research areas

Related content

IN, TS, Hyderabad
Welcome to the Worldwide Returns & ReCommerce team (WWR&R) at Amazon.com. WWR&R is an agile, innovative organization dedicated to ‘making zero happen’ to benefit our customers, our company, and the environment. Our goal is to achieve the three zeroes: zero cost of returns, zero waste, and zero defects. We do this by developing products and driving truly innovative operational excellence to help customers keep what they buy, recover returned and damaged product value, keep thousands of tons of waste from landfills, and create the best customer returns experience in the world. We have an eye to the future – we create long-term value at Amazon by focusing not just on the bottom line, but on the planet. We are building the most sustainable re-use channel we can by driving multiple aspects of the Circular Economy for Amazon – Returns & ReCommerce. Amazon WWR&R is comprised of business, product, operational, program, software engineering and data teams that manage the life of a returned or damaged product from a customer to the warehouse and on to its next best use. Our work is broad and deep: we train machine learning models to automate routing and find signals to optimize re-use; we invent new channels to give products a second life; we develop highly respected product support to help customers love what they buy; we pilot smarter product evaluations; we work from the customer backward to find ways to make the return experience remarkably delightful and easy; and we do it all while scrutinizing our business with laser focus. You will help create everything from customer-facing and vendor-facing websites to the internal software and tools behind the reverse-logistics process. You can develop scalable, high-availability solutions to solve complex and broad business problems. We are a group that has fun at work while driving incredible customer, business, and environmental impact. We are backed by a strong leadership group dedicated to operational excellence that empowers a reasonable work-life balance. As an established, experienced team, we offer the scope and support needed for substantial career growth. Amazon is earth’s most customer-centric company and through WWR&R, the earth is our customer too. Come join us and innovate with the Amazon Worldwide Returns & ReCommerce team!
GB, MLN, Edinburgh
We’re looking for a Machine Learning Scientist in the Personalization team for our Edinburgh office experienced in generative AI and large models. You will be responsible for developing and disseminating customer-facing personalized recommendation models. This is a hands-on role with global impact working with a team of world-class engineers and scientists across the Edinburgh offices and wider organization. You will lead the design of machine learning models that scale to very large quantities of data, and serve high-scale low-latency recommendations to all customers worldwide. You will embody scientific rigor, designing and executing experiments to demonstrate the technical efficacy and business value of your methods. You will work alongside a science team to delight customers by aiding in recommendations relevancy, and raise the profile of Amazon as a global leader in machine learning and personalization. Successful candidates will have strong technical ability, focus on customers by applying a customer-first approach, excellent teamwork and communication skills, and a motivation to achieve results in a fast-paced environment. Our position offers exceptional opportunities for every candidate to grow their technical and non-technical skills. If you are selected, you have the opportunity to make a difference to our business by designing and building state of the art machine learning systems on big data, leveraging Amazon’s vast computing resources (AWS), working on exciting and challenging projects, and delivering meaningful results to customers world-wide. Key job responsibilities Develop machine learning algorithms for high-scale recommendations problems. Rapidly design, prototype and test many possible hypotheses in a high-ambiguity environment, making use of both quantitative analysis and business judgement. Collaborate with software engineers to integrate successful experimental results into large-scale, highly complex Amazon production systems capable of handling 100,000s of transactions per second at low latency. Report results in a manner which is both statistically rigorous and compellingly relevant, exemplifying good scientific practice in a business environment.
US, WA, Seattle
Prime Video is a first-stop entertainment destination offering customers a vast collection of premium programming in one app available across thousands of devices. Prime members can customize their viewing experience and find their favorite movies, series, documentaries, and live sports – including Amazon MGM Studios-produced series and movies; licensed fan favorites; and programming from Prime Video add-on subscriptions such as Apple TV+, Max, Crunchyroll and MGM+. All customers, regardless of whether they have a Prime membership or not, can rent or buy titles via the Prime Video Store, and can enjoy even more content for free with ads. Are you interested in shaping the future of entertainment? Prime Video's technology teams are creating best-in-class digital video experience. As a Prime Video technologist, you’ll have end-to-end ownership of the product, user experience, design, and technology required to deliver state-of-the-art experiences for our customers. You’ll get to work on projects that are fast-paced, challenging, and varied. You’ll also be able to experiment with new possibilities, take risks, and collaborate with remarkable people. We’ll look for you to bring your diverse perspectives, ideas, and skill-sets to make Prime Video even better for our customers. With global opportunities for talented technologists, you can decide where a career Prime Video Tech takes you! In Prime Video READI, our mission is to automate infrastructure scaling and operational readiness. We are growing a team specialized in time series modeling, forecasting, and release safety. This team will invent and develop algorithms for forecasting multi-dimensional related time series. The team will develop forecasts on key business dimensions with optimization recommendations related to performance and efficiency opportunities across our global software environment. As a founding member of the core team, you will apply your deep coding, modeling and statistical knowledge to concrete problems that have broad cross-organizational, global, and technology impact. Your work will focus on retrieving, cleansing and preparing large scale datasets, training and evaluating models and deploying them to production where we continuously monitor and evaluate. You will work on large engineering efforts that solve significantly complex problems facing global customers. You will be trusted to operate with complete independence and are often assigned to focus on areas where the business and/or architectural strategy has not yet been defined. You must be equally comfortable digging in to business requirements as you are drilling into design with development teams and developing production ready learning models. You consistently bring strong, data-driven business and technical judgment to decisions. You will work with internal and external stakeholders, cross-functional partners, and end-users around the world at all levels. Our team makes a big impact because nothing is more important to us than delivering for our customers, continually earning their trust, and thinking long term. You are empowered to bring new technologies to your solutions. If you crave a sense of ownership, this is the place to be.
US, WA, Seattle
Amazon Advertising operates at the intersection of eCommerce and advertising, and is investing heavily in building a world-class advertising business. We are defining and delivering a collection of self-service performance advertising products that drive discovery and sales. Our products are strategically important to our Retail and Marketplace businesses driving long-term growth. We deliver billions of ad impressions and millions of clicks daily and are breaking fresh ground to create world-class products to improve both shopper and advertiser experience. With a broad mandate to experiment and innovate, we grow at an unprecedented rate with a seemingly endless range of new opportunities. The Ad Response Prediction team in Sponsored Products organization build advanced deep-learning models, large-scale machine-learning pipelines, and real-time serving infra to match shoppers’ intent to relevant ads on all devices, for all contexts and in all marketplaces. Through precise estimation of shoppers’ interaction with ads and their long-term value, we aim to drive optimal ads allocation and pricing, and help to deliver a relevant, engaging and delightful ads experience to Amazon shoppers. As the business and the complexity of various new initiatives we take continues to grow, we are looking for talented Applied Scientists to join the team. Key job responsibilities As a Applied Scientist II, you will: * Conduct hands-on data analysis, build large-scale machine-learning models and pipelines * Work closely with software engineers on detailed requirements, technical designs and implementation of end-to-end solutions in production * Run regular A/B experiments, gather data, perform statistical analysis, and communicate the impact to senior management * Establish scalable, efficient, automated processes for large-scale data analysis, machine-learning model development, model validation and serving * Provide technical leadership, research new machine learning approaches to drive continued scientific innovation * Be a member of the Amazon-wide Machine Learning Community, participating in internal and external MeetUps, Hackathons and Conferences
US, WA, Seattle
Prime Video is a first-stop entertainment destination offering customers a vast collection of premium programming in one app available across thousands of devices. Prime members can customize their viewing experience and find their favorite movies, series, documentaries, and live sports – including Amazon MGM Studios-produced series and movies; licensed fan favorites; and programming from Prime Video add-on subscriptions such as Apple TV+, Max, Crunchyroll and MGM+. All customers, regardless of whether they have a Prime membership or not, can rent or buy titles via the Prime Video Store, and can enjoy even more content for free with ads. Are you interested in shaping the future of entertainment? Prime Video's technology teams are creating best-in-class digital video experience. As a Prime Video team member, you’ll have end-to-end ownership of the product, user experience, design, and technology required to deliver state-of-the-art experiences for our customers. You’ll get to work on projects that are fast-paced, challenging, and varied. You’ll also be able to experiment with new possibilities, take risks, and collaborate with remarkable people. We’ll look for you to bring your diverse perspectives, ideas, and skill-sets to make Prime Video even better for our customers. With global opportunities for talented technologists, you can decide where a career Prime Video Tech takes you! Key job responsibilities As an Applied Scientist in the Content Understanding Team, you will lead the end-to-end research and deployment of video and multi-modal models applied to a variety of downstream applications. More specifically, you will: - Work backwards from customer problems to research and design scientific approaches for solving them - Work closely with other scientists, engineers and product managers to expand the depth of our product insights with data, create a variety of experiments to determine the high impact projects to include in planning roadmaps - Stay up-to-date with advancements and the latest modeling techniques in the field - Publish your research findings in top conferences and journals About the team Our Prime Video Content Understanding team builds holistic media representations (e.g. descriptions of scenes, semantic embeddings) and apply them to new customer experiences supply chain problems. Our technology spans the entire Prime Video catalogue globally, and we enable instant recaps, skip intro timing, ad placement, search, and content moderation.
IN, HR, Gurugram
We're on a journey to build something new a green field project! Come join our team and build new discovery and shopping products that connect customers with their vehicle of choice. We're looking for a talented Senior Applied Scientist to join our team of product managers, designers, and engineers to design, and build innovative automotive-shopping experiences for our customers. This is a great opportunity for an experienced engineer to design and implement the technology for a new Amazon business. We are looking for a Applied Scientist to design, implement and deliver end-to-end solutions. We are seeking passionate, hands-on, experienced and seasoned Senior Applied Scientist who will be deep in code and algorithms; who are technically strong in building scalable computer vision machine learning systems across item understanding, pose estimation, class imbalanced classifiers, identification and segmentation.. You will drive ideas to products using paradigms such as deep learning, semi supervised learning and dynamic learning. As a Senior Applied Scientist, you will also help lead and mentor our team of applied scientists and engineers. You will take on complex customer problems, distill customer requirements, and then deliver solutions that either leverage existing academic and industrial research or utilize your own out-of-the-box but pragmatic thinking. In addition to coming up with novel solutions and prototypes, you will directly contribute to implementation while you lead. A successful candidate has excellent technical depth, scientific vision, project management skills, great communication skills, and a drive to achieve results in a unified team environment. You should enjoy the process of solving real-world problems that, quite frankly, haven’t been solved at scale anywhere before. Along the way, we guarantee you’ll get opportunities to be a bold disruptor, prolific innovator, and a reputed problem solver—someone who truly enables AI and robotics to significantly impact the lives of millions of consumers. Key job responsibilities Architect, design, and implement Machine Learning models for vision systems on robotic platforms Optimize, deploy, and support at scale ML models on the edge. Influence the team's strategy and contribute to long-term vision and roadmap. Work with stakeholders across , science, and operations teams to iterate on design and implementation. Maintain high standards by participating in reviews, designing for fault tolerance and operational excellence, and creating mechanisms for continuous improvement. Prototype and test concepts or features, both through simulation and emulators and with live robotic equipment Work directly with customers and partners to test prototypes and incorporate feedback Mentor other engineer team members. A day in the life - 6+ years of building machine learning models for retail application experience - PhD, or Master's degree and 6+ years of applied research experience - Experience programming in Java, C++, Python or related language - Experience with neural deep learning methods and machine learning - Demonstrated expertise in computer vision and machine learning techniques.
US, WA, Seattle
Prime Video is a first-stop entertainment destination offering customers a vast collection of premium programming in one app available across thousands of devices. Prime members can customize their viewing experience and find their favorite movies, series, documentaries, and live sports – including Amazon MGM Studios-produced series and movies; licensed fan favorites; and programming from Prime Video add-on subscriptions such as Apple TV+, Max, Crunchyroll and MGM+. All customers, regardless of whether they have a Prime membership or not, can rent or buy titles via the Prime Video Store, and can enjoy even more content for free with ads. Are you interested in shaping the future of entertainment? Prime Video's technology teams are creating best-in-class digital video experience. As a Prime Video team member, you’ll have end-to-end ownership of the product, user experience, design, and technology required to deliver state-of-the-art experiences for our customers. You’ll get to work on projects that are fast-paced, challenging, and varied. You’ll also be able to experiment with new possibilities, take risks, and collaborate with remarkable people. We’ll look for you to bring your diverse perspectives, ideas, and skill-sets to make Prime Video even better for our customers. With global opportunities for talented technologists, you can decide where a career Prime Video Tech takes you! Key job responsibilities As an Applied Scientist in the Content Understanding Team, you will lead the end-to-end research and deployment of video and multi-modal models applied to a variety of downstream applications. More specifically, you will: - Work backwards from customer problems to research and design scientific approaches for solving them - Work closely with other scientists, engineers and product managers to expand the depth of our product insights with data, create a variety of experiments to determine the high impact projects to include in planning roadmaps - Stay up-to-date with advancements and the latest modeling techniques in the field - Publish your research findings in top conferences and journals About the team Our Prime Video Content Understanding team builds holistic media representations (e.g. descriptions of scenes, semantic embeddings) and apply them to new customer experiences supply chain problems. Our technology spans the entire Prime Video catalogue globally, and we enable instant recaps, skip intro timing, ad placement, search, and content moderation.
US, WA, Seattle
Do you want to re-invent how millions of people consume video content on their TVs, Tablets and Alexa? We are building a free to watch streaming service called Fire TV Channels (https://techcrunch.com/2023/08/21/amazon-launches-fire-tv-channels-app-400-fast-channels/). Our goal is to provide customers with a delightful and personalized experience for consuming content across News, Sports, Cooking, Gaming, Entertainment, Lifestyle and more. You will work closely with engineering and product stakeholders to realize our ambitious product vision. You will get to work with Generative AI and other state of the art technologies to help build personalization and recommendation solutions from the ground up. You will be in the driver's seat to present customers with content they will love. Using Amazon’s large-scale computing resources, you will ask research questions about customer behavior, build state-of-the-art models to generate recommendations and run these models to enhance the customer experience. You will participate in the Amazon ML community and mentor Applied Scientists and Software Engineers with a strong interest in and knowledge of ML. Your work will directly benefit customers and you will measure the impact using scientific tools.
US, MA, Boston
The Artificial General Intelligence (AGI) team is looking for a passionate, talented, and inventive Applied Scientist with a strong deep learning background, to build industry-leading technology with Large Language Models (LLMs) and multi-modal systems. You will support projects that work on technologies including multi-modal model alignment, moderation systems and evaluation. Key job responsibilities As an Applied Scientist with the AGI team, you will support the development of novel algorithms and modeling techniques, to advance the state of the art with LLMs. Your work will directly impact our customers in the form of products and services that make use of speech and language technology. You will leverage Amazon’s heterogeneous data sources and large-scale computing resources to accelerate advances in generative artificial intelligence (GenAI). You are also expected to publish in top tier conferences. About the team The AGI team has a mission to push the envelope in LLMs and multimodal systems. Specifically, we focus on model alignment with an aim to maintain safety while not denting utility, in order to provide the best-possible experience for our customers.
IN, HR, Gurugram
Our customers have immense faith in our ability to deliver packages timely and as expected. A well planned network seamlessly scales to handle millions of package movements a day. It has monitoring mechanisms that detect failures before they even happen (such as predicting network congestion, operations breakdown), and perform proactive corrective actions. When failures do happen, it has inbuilt redundancies to mitigate impact (such as determine other routes or service providers that can handle the extra load), and avoids relying on single points of failure (service provider, node, or arc). Finally, it is cost optimal, so that customers can be passed the benefit from an efficiently set up network. Amazon Shipping is hiring Applied Scientists to help improve our ability to plan and execute package movements. As an Applied Scientist in Amazon Shipping, you will work on multiple challenging machine learning problems spread across a wide spectrum of business problems. You will build ML models to help our transportation cost auditing platforms effectively audit off-manifest (discrepancies between planned and actual shipping cost). You will build models to improve the quality of financial and planning data by accurately predicting ship cost at a package level. Your models will help forecast the packages required to be pick from shipper warehouses to reduce First Mile shipping cost. Using signals from within the transportation network (such as network load, and velocity of movements derived from package scan events) and outside (such as weather signals), you will build models that predict delivery delay for every package. These models will help improve buyer experience by triggering early corrective actions, and generating proactive customer notifications. Your role will require you to demonstrate Think Big and Invent and Simplify, by refining and translating Transportation domain-related business problems into one or more Machine Learning problems. You will use techniques from a wide array of machine learning paradigms, such as supervised, unsupervised, semi-supervised and reinforcement learning. Your model choices will include, but not be limited to, linear/logistic models, tree based models, deep learning models, ensemble models, and Q-learning models. You will use techniques such as LIME and SHAP to make your models interpretable for your customers. You will employ a family of reusable modelling solutions to ensure that your ML solution scales across multiple regions (such as North America, Europe, Asia) and package movement types (such as small parcel movements and truck movements). You will partner with Applied Scientists and Research Scientists from other teams in US and India working on related business domains. Your models are expected to be of production quality, and will be directly used in production services. You will work as part of a diverse data science and engineering team comprising of other Applied Scientists, Software Development Engineers and Business Intelligence Engineers. You will participate in the Amazon ML community by authoring scientific papers and submitting them to Machine Learning conferences. You will mentor Applied Scientists and Software Development Engineers having a strong interest in ML. You will also be called upon to provide ML consultation outside your team for other problem statements. If you are excited by this charter, come join us!