Bruno Zamborlin (born 1983 in Vicenza) is an AI researcher, entrepreneur and artist based in London, working in the field of human-computer interaction. His work focuses on converting physical objects into touch-sensitive, interactive surfaces using vibration sensors and artificial intelligence. In 2013, he founded Mogees Limited a start-up to transform everyday objects into musical instruments and games using a vibration sensor and a mobile phone. With HyperSurfaces, he converts physical surfaces of any material, shape and form into data-enabled-interactive surfaces using a vibration sensor and a coin-sized chipset. As an artist, he has created art installations around the world, with his most recent work comprising a unique series of "sound furnitures" that was showcased at the Italian Pavilion of the Venice Biennale 2023. He regularly performed with UK-based electronic music duo Plaid (Warp Records). He is also honorary visiting research fellow at Goldsmiths, University of London. == Early life and education == From 2008-2011, Zamborlin worked at the IRCAM (Institute for Research and Coordination Acoustic Musical) – Centre Pompidou as a member of the Sound Music Movement Interaction team. Under the supervision of Frederic Bevilacqua, he started experimenting with the use of artificial intelligence and human movements, and contributed to the creation of Gesture Follower, a software used to analyse body movements of performers and dancers through motion sensors in order to control sound and visual media in real-time, slowing down or speeding up their reproduction based on the speed the gestures are performed. He has lived in London since 2011, where he developed a joint PhD between Goldsmiths, University of London and IRCAM - Centre Pompidou/Pierre and Marie Curie University Paris in AI, focussing on the concept of Interactive Machine Learning applied to digital musical instruments and performing arts. == Career == Zamborlin founded Mogees Limited in 2013 in London, with IRCAM being amongst the early partners. Mogees transform physical objects into musical instruments and games using a vibration sensor and a series of apps for smartphones and desktop. After a campaign on Kickstarter in 2014, Mogees was used both by common users and artists such as Rodrigo y Gabriela, Jean-Michel Jarre and Plaid. The algorithms implemented in these apps employ a special version of physical modelling sound synthesis, where the vibration produced by users when interacting with the physical object are used as exciter for a digital resonator which runs in the app. The result is a hybrid, half acoustic and half digital sound which is a function of both software and acoustic properties of the physical object the users decide to play. In 2017, Zamborlin founded HyperSurfaces together with computational artist Parag K Mital. to merge "the physical and the digital worlds". HyperSurfaces technology converts any surface made of any material, shape and size into data-enabled interactive objects, employing a vibration sensor and proprietary AI algorithms running on a coin-sized chipset. The vibrations generated by people's interactions on the surface are converted into an electric signal by a piezoelectric sensor and analysed in realtime by AI algorithms that run on the chipset. Anytime the AI recognises in the vibration signal one of the events that have been predefined by the user beforehand, a corresponding notification message is generated in realtime and sent to some application. The technology can be applied to anything ranging from button-less human-computer interaction applications for automotive and smart home to the Internet of things. Because the AI algorithms employed by HyperSurfaces run locally on a chipset, without the need to access cloud-based services, they are considered to be part of the field of edge computing. Also, because the AI can be trained beforehand to recognise the events its users are interested in, HyperSurfaces algorithms belong to the field of supervised machine learning. == Selected awards == IRISA Prix Jeune Chercheur, 13 October 2012 NeMoDe, New Economic Models in the Digital Economy, 25 October 2012 == Patents and academic publications == United States pending US10817798B2, Bruno Zamborlin & Carmine Emanuele Cella, "Method to recognize a gesture and corresponding device", published 27 April 2016, assigned to Mogees Limited GB Pending WO/2019/086862, Bruno Zamborlin; Conor Barry & Alessandro Saccoia et al., "A user interface for vehicles", published 9 May 2019, assigned to Mogees Limited GB Pending WO/2019/086863, Bruno Zamborlin; Conor Barry & Alessandro Saccoia et al., "Trigger for game events", published 9 May 2019, assigned to Mogees Limited Bevilacqua, Frédéric; Zamborlin, Bruno; Sypniewski, Anthony; Schnell, Norbert; Guédy, Fabrice; Rasamimanana, Nicolas (2010). "Continuous Realtime Gesture Following and Recognition". Gesture in Embodied Communication and Human-Computer Interaction. Lecture Notes in Computer Science. Vol. 5934. pp. 73–84. doi:10.1007/978-3-642-12553-9_7. ISBN 978-3-642-12552-2. S2CID 16251822. Retrieved 17 January 2021. Rasamimanana, Nicolas; Bevilacqua, Frédéric; Schnell, Norbert; Guédy, Fabrice; Flety, Emmanuel; Maestracci, Come; Zamborlin, Bruno (January 2010). "Modular musical objects towards embodied control of digital music". Proceedings of the fifth international conference on Tangible, embedded, and embodied interaction. Tei '11. pp. 9–12. doi:10.1145/1935701.1935704. ISBN 9781450304788. S2CID 10782645. Retrieved 17 January 2021. Bevilacqua, Frédéric; Schnell, Norbert; Rasamimanana, Nicolas; Zamborlin, Bruno; Guedy, Fabrice (2011). "Online Gesture Analysis and Control of Audio Processing". Musical Robots and Interactive Multimodal Systems. Springer Tracts in Advanced Robotics. Vol. 74. pp. 127–142. doi:10.1007/978-3-642-22291-7_8. ISBN 978-3-642-22290-0. Retrieved 17 January 2021. Zamborlin, Bruno; Bevilacqua, Frédéric; Gillies, Marco; D'Inverno, Mark (15 January 2014). "Fluid gesture interaction design: Applications of continuous recognition for the design of modern gestural interfaces". ACM Transactions on Interactive Intelligent Systems. 3 (4): 22:1–22:30. doi:10.1145/2543921. S2CID 7887245. Retrieved 17 January 2021. Leslie, Grace; Zamborlin, Bruno; Schnell, Norbert; Jodlowski, Pierre (15 June 2010). "A Collaborative, Interactive Sound Installation". Proceedings of the International Computer Music Conference. Retrieved 17 January 2021. Kimura, Mari; Rasamimanana, Nicolas; Bevilacqua, Frédéric; Zamborlin, Bruno; Schnell, Bruno; Flety, Emmanuel (2012). "Extracting Human Expression For Interactive Composition with the Augmented Violin". International Conference on New Interfaces for Musical Expression. Retrieved 17 January 2021. Ferretti, Stefano; Roccetti, Marco; Zamborlin, Bruno (13 January 2009). "On SPAWC: Discussion on a Musical Signal Parser and Well-Formed Composer". 2009 6th IEEE Consumer Communications and Networking Conference. pp. 1–5. doi:10.1109/CCNC.2009.4784966. ISBN 978-1-4244-2308-8. S2CID 14213587. Zamborlin, Bruno; Partesana, Giorgio; Liuni, Marco (15 May 2011). "(LAND)MOVES". Conference on New Interfaces for Musical Expression, NIME: 537–538. Retrieved 17 January 2021.
Spike (application)
Spike is a cross-platform email client and AI-powered communication app, available on Windows, MacOS, iOS, Android and the web. It has a chat-like, conversational view for emails with AI-powered inbox management and integrated collaboration features. Depending on the selected plan, it can be used solely as an email application or as a full suite of business communication tools. == History == Founded in 2013 by Erez Pilosof and Dvir Ben-Aroya, Spike is a software application that puts existing e-mails into a multimedia messaging, chat-like interface enhanced with video and voice calls. The application was initially named Hop. In 2019, the developers completed a $5 million funding round including investment from Wix.com and NFX Capital. In 2020, Spike raised $8m in a Series A funding round led by Insight Partners with the participation from previous rounds' investors. In 2021 Spike announced a collaboration with Meta to launch on the Oculus Store and would become one of the first productivity apps to launch in Meta's new virtual world, known as the Metaverse. In June 2023, the company introduced its corporate offering — Teamspace, a corporate communication platform for teams with features such as company-wide channels for broad conversations, private groups for specific topics or projects, direct one-on-one conversations, video meetings, file collaboration, AI-powered email messaging, and custom email domain. It supports file management, search capabilities, and project management. Built on open-protocol technology, Spike Teamspace enables users to send and receive messages from all email providers. Regardless of whether the other party is using Spike. == Company operations == Spike is developed and operated by SpikeNow LTD. Dvir Ben Aroya serves as Spike’s CEO and Erez Pilosof is the CTO. The company is headquartered in Tel Aviv, Israel. == Mode of use == The app enables users to organize email into three types of "conversations,"a traditional inbox/sent format, by subject, or by people. Spike users can also make audio and video calls to each other, and other features include a calendar, contact list, and Groups. Spike is available for Microsoft Windows, MacOS, iOS and Android, and as a web version, and works with Gmail, Outlook, Exchange, iCloud, Yahoo! Mail and IMAP email providers. == Features == Since 2023, the platform features an AI-driven assistant, Magic AI, for customized email creation, document summarization, research, content generation, advanced note-taking, project management, and real-time translation. Since 2023, Spike offers custom email domain management. It supports team collaboration through Channels, uniting members globally with access to historical messages, and combines email with real-time messaging via Conversational Email. The Shared Inbox allows team collaboration on emails, while Groups support private conversations and invitations. It also features integrated video meetings, real-time collaboration on documents and notes, and email hosting with custom domains. Super Search enables retrieval of various content, and the Priority Inbox organizes emails by priority. Collaborative Tasks offer real-time updates and tracking. The platform allows voice message sending from mobile devices and integrates multiple calendar platforms into a unified schedule. File Management optimizes attachment handling, and the Unified Inbox consolidates emails from multiple accounts. Spike ensures data security with AES-256 encryption and private keys. The platform features AI-powered inbox management and communication tools. In May 2025, Spike launched its AI Feed feature, which automatically summarizes unread messages in a unified stream and enables bulk email actions. Additional AI capabilities include email composition assistance, document summarization, content generation, note-taking enhancement, and real-time translation.
Structured kNN
Structured k-nearest neighbours (SkNN) is a machine learning algorithm that generalizes k-nearest neighbors (k-NN). k-NN supports binary classification, multiclass classification, and regression, whereas SkNN allows training of a classifier for general structured output. For instance, a data sample might be a natural language sentence, and the output could be an annotated parse tree. Training a classifier consists of showing many instances of ground truth sample-output pairs. After training, the SkNN model is able to predict the corresponding output for new, unseen sample instances; that is, given a natural language sentence, the classifier can produce the most likely parse tree. == Training == As a training set, SkNN accepts sequences of elements with class labels. The type of element does not matter; the only requirement is a defined metric function that gives a distance between each pair of elements of a set. SkNN is based on idea of creating a graph, with each node representing a class label. There is an edge between a pair of nodes if there is a sequence of two elements in the training set with corresponding classes. The first step of SkNN training is the construction of such a graph from training sequences. There are two special nodes in the graph corresponding to sentence beginnings and ends: if a sequence starts with class C, the edge between node START and node C should be created. Like regular k-NN, the second part of SkNN training consists of storing the elements of a training sequence in a certain way. Each element of the training sequences is stored in the node related to the class of the previous element in the sequence. Every first element is stored in the START node. == Inference == Labelling input sequences by SkNN consists of finding the sequence of transitions in the graph, starting from node START. Each transition corresponds to a single element of the input sequence. As a result, the label of each element is determined as the target node label of the transition. The cost of the path is defined as the sum of all transitions, with the cost of transition from node A to node B being the distance from the current input sequence element to the nearest element of class B, stored in node A. Determining an optimal path may be performed using a modified Viterbi algorithm (where the sum of the distances is minimized, unlike the original algorithm which maximizes the product of probabilities).
Prototype methods
Prototype methods are machine learning methods that use data prototypes. A data prototype is a data value that reflects other values in its class, e.g., the centroid in a K-means clustering problem. == Methods == The following are some prototype methods K-means clustering Learning vector quantization (LVQ) Gaussian mixtures == Related Methods == While K-nearest neighbor's does not use prototypes, it is similar to prototype methods like K-means clustering.
Proximal policy optimization
Proximal policy optimization (PPO) is a reinforcement learning (RL) algorithm for training an intelligent agent. Specifically, it is a policy gradient method, often used for deep RL when the policy network is very large. == History == The predecessor to PPO, Trust Region Policy Optimization (TRPO), was published in 2015. It addressed the instability issue of another algorithm, the Deep Q-Network (DQN), by using the trust region method to limit the KL divergence between the old and new policies. However, TRPO uses the Hessian matrix (a matrix of second derivatives) to enforce the trust region, but the Hessian is inefficient for large-scale problems. PPO was published in 2017. It was essentially an approximation of TRPO that does not require computing the Hessian. The KL divergence constraint was approximated by simply clipping the policy gradient. Since 2018, PPO was the default RL algorithm at OpenAI. PPO has been applied to many areas, such as controlling a robotic arm, beating professional players at Dota 2 (OpenAI Five), and playing Atari games. == TRPO == TRPO, the predecessor of PPO, is an on-policy algorithm. It can be used for environments with either discrete or continuous action spaces. The pseudocode is as follows: Input: initial policy parameters θ 0 {\textstyle \theta _{0}} , initial value function parameters ϕ 0 {\textstyle \phi _{0}} Hyperparameters: KL-divergence limit δ {\textstyle \delta } , backtracking coefficient α {\textstyle \alpha } , maximum number of backtracking steps K {\textstyle K} for k = 0 , 1 , 2 , … {\textstyle k=0,1,2,\ldots } do Collect set of trajectories D k = { τ i } {\textstyle {\mathcal {D}}_{k}=\left\{\tau _{i}\right\}} by running policy π k = π ( θ k ) {\textstyle \pi _{k}=\pi \left(\theta _{k}\right)} in the environment. Compute rewards-to-go R ^ t {\textstyle {\hat {R}}_{t}} . Compute advantage estimates, A ^ t {\textstyle {\hat {A}}_{t}} (using any method of advantage estimation) based on the current value function V ϕ k {\textstyle V_{\phi _{k}}} . Estimate policy gradient as g ^ k = 1 | D k | ∑ τ ∈ D k ∑ t = 0 T ∇ θ log π θ ( a t ∣ s t ) | θ k A ^ t {\displaystyle {\hat {g}}_{k}=\left.{\frac {1}{\left|{\mathcal {D}}_{k}\right|}}\sum _{\tau \in {\mathcal {D}}_{k}}\sum _{t=0}^{T}\nabla _{\theta }\log \pi _{\theta }\left(a_{t}\mid s_{t}\right)\right|_{\theta _{k}}{\hat {A}}_{t}} Use the conjugate gradient algorithm to compute x ^ k ≈ H ^ k − 1 g ^ k {\displaystyle {\hat {x}}_{k}\approx {\hat {H}}_{k}^{-1}{\hat {g}}_{k}} where H ^ k {\textstyle {\hat {H}}_{k}} is the Hessian of the sample average KL-divergence. Update the policy by backtracking line search with θ k + 1 = θ k + α j 2 δ x ^ k T H ^ k x ^ k x ^ k {\displaystyle \theta _{k+1}=\theta _{k}+\alpha ^{j}{\sqrt {\frac {2\delta }{{\hat {x}}_{k}^{T}{\hat {H}}_{k}{\hat {x}}_{k}}}}{\hat {x}}_{k}} where j ∈ { 0 , 1 , 2 , … K } {\textstyle j\in \{0,1,2,\ldots K\}} is the smallest value which improves the sample loss and satisfies the sample KL-divergence constraint. Fit value function by regression on mean-squared error: ϕ k + 1 = arg min ϕ 1 | D k | T ∑ τ ∈ D k ∑ t = 0 T ( V ϕ ( s t ) − R ^ t ) 2 {\displaystyle \phi _{k+1}=\arg \min _{\phi }{\frac {1}{\left|{\mathcal {D}}_{k}\right|T}}\sum _{\tau \in {\mathcal {D}}_{k}}\sum _{t=0}^{T}\left(V_{\phi }\left(s_{t}\right)-{\hat {R}}_{t}\right)^{2}} typically via some gradient descent algorithm. == PPO == The pseudocode is as follows: Input: initial policy parameters θ 0 {\textstyle \theta _{0}} , initial value function parameters ϕ 0 {\textstyle \phi _{0}} for k = 0 , 1 , 2 , … {\textstyle k=0,1,2,\ldots } do Collect set of trajectories D k = { τ i } {\textstyle {\mathcal {D}}_{k}=\left\{\tau _{i}\right\}} by running policy π k = π ( θ k ) {\textstyle \pi _{k}=\pi \left(\theta _{k}\right)} in the environment. Compute rewards-to-go R ^ t {\textstyle {\hat {R}}_{t}} . Compute advantage estimates, A ^ t {\textstyle {\hat {A}}_{t}} (using any method of advantage estimation) based on the current value function V ϕ k {\textstyle V_{\phi _{k}}} . Update the policy by maximizing the PPO-Clip objective: θ k + 1 = arg max θ 1 | D k | T ∑ τ ∈ D k ∑ t = 0 T min ( π θ ( a t ∣ s t ) π θ k ( a t ∣ s t ) A π θ k ( s t , a t ) , g ( ϵ , A π θ k ( s t , a t ) ) ) {\displaystyle \theta _{k+1}=\arg \max _{\theta }{\frac {1}{\left|{\mathcal {D}}_{k}\right|T}}\sum _{\tau \in {\mathcal {D}}_{k}}\sum _{t=0}^{T}\min \left({\frac {\pi _{\theta }\left(a_{t}\mid s_{t}\right)}{\pi _{\theta _{k}}\left(a_{t}\mid s_{t}\right)}}A^{\pi _{\theta _{k}}}\left(s_{t},a_{t}\right),\quad g\left(\epsilon ,A^{\pi _{\theta _{k}}}\left(s_{t},a_{t}\right)\right)\right)} typically via stochastic gradient ascent with Adam. Fit value function by regression on mean-squared error: ϕ k + 1 = arg min ϕ 1 | D k | T ∑ τ ∈ D k ∑ t = 0 T ( V ϕ ( s t ) − R ^ t ) 2 {\displaystyle \phi _{k+1}=\arg \min _{\phi }{\frac {1}{\left|{\mathcal {D}}_{k}\right|T}}\sum _{\tau \in {\mathcal {D}}_{k}}\sum _{t=0}^{T}\left(V_{\phi }\left(s_{t}\right)-{\hat {R}}_{t}\right)^{2}} typically via some gradient descent algorithm. Like all policy gradient methods, PPO is used for training an RL agent whose actions are determined by a differentiable policy function by gradient ascent. Intuitively, a policy gradient method takes small policy update steps, so the agent can reach higher and higher rewards in expectation. Policy gradient methods may be unstable: A step size that is too big may direct the policy in a suboptimal direction, thus having little possibility of recovery; a step size that is too small lowers the overall efficiency. To solve the instability, PPO implements a clip function that constrains the policy update of an agent from being too large, so that larger step sizes may be used without negatively affecting the gradient ascent process. === Basic concepts === To begin the PPO training process, the agent is set in an environment to perform actions based on its current input. In the early phase of training, the agent can freely explore solutions and keep track of the result. Later, with a certain amount of transition samples and policy updates, the agent will select an action to take by randomly sampling from the probability distribution P ( A | S ) {\displaystyle P(A|S)} generated by the policy network. The actions that are most likely to be beneficial will have the highest probability of being selected from the random sample. After an agent arrives at a different scenario (a new state) by acting, it is rewarded with a positive reward or a negative reward. The objective of an agent is to maximize the cumulative reward signal across sequences of states, known as episodes. === Policy gradient laws: the advantage function === The advantage function (denoted as A {\displaystyle A} ) is central to PPO, as it tries to answer the question of whether a specific action of the agent is better or worse than some other possible action in a given state. By definition, the advantage function is an estimate of the relative value for a selected action. If the output of this function is positive, it means that the action in question is better than the average return, so the possibilities of selecting that specific action will increase. The opposite is true for a negative advantage output. The advantage function can be defined as A = Q − V {\displaystyle A=Q-V} , where Q {\displaystyle Q} is the discounted sum of rewards (the total weighted reward for the completion of an episode) and V {\displaystyle V} is the baseline estimate. Since the advantage function is calculated after the completion of an episode, the program records the outcome of the episode. Therefore, calculating advantage is essentially an unsupervised learning problem. The baseline estimate comes from the value function that outputs the expected discounted sum of an episode starting from the current state. In the PPO algorithm, the baseline estimate will be noisy (with some variance), as it also uses a neural network, like the policy function itself. With Q {\displaystyle Q} and V {\displaystyle V} computed, the advantage function is calculated by subtracting the baseline estimate from the actual discounted return. If A > 0 {\displaystyle A>0} , the actual return of the action is better than the expected return from experience; if A < 0 {\displaystyle A<0} , the actual return is worse. === Ratio function === In PPO, the ratio function ( r t {\displaystyle r_{t}} ) calculates the probability of selecting action a {\displaystyle a} in state s {\displaystyle s} given the current policy network, divided by the previous probability under the old policy. In other words: If r t ( θ ) > 1 {\displaystyle r_{t}(\theta )>1} , where θ {\displaystyle \theta } are the policy network parameters, then selecting action a {\displaystyle a} in state s {\displaystyle s} is more likely based on the current policy than the previous policy. If 0 ≤ r t ( θ ) < 1 {\displaystyle 0\leq r_{t}(\theta )<1} , then selecting actio
Reasoning model
A reasoning model, also known as a reasoning language model (RLM) or large reasoning model (LRM), is a type of large language model (LLM) that has been specifically trained to solve complex tasks requiring multiple steps of logical reasoning. These models demonstrate superior performance on logic, mathematics, and programming tasks compared to standard LLMs. They possess the ability to revisit and revise earlier reasoning steps and utilize additional computation during inference as a method to scale performance, complementing traditional scaling approaches based on training data size, model parameters, and training compute. == Overview == Unlike traditional language models that generate responses immediately, reasoning models allocate additional compute, or thinking, time before producing an answer to solve multi-step problems. OpenAI introduced this terminology in September 2024 when it released the o1 series, describing the models as designed to "spend more time thinking" before responding. The company framed o1 as a reset in model naming that targets complex tasks in science, coding, and mathematics, and it contrasted o1's performance with GPT-4o on benchmarks such as AIME and Codeforces. Independent reporting the same week summarized the launch and highlighted OpenAI's claim that o1 automates chain-of-thought style reasoning to achieve large gains on difficult exams. In operation, reasoning models generate internal chains of intermediate steps, then select and refine a final answer. OpenAI reported that o1's accuracy improves as the model is given more reinforcement learning during training and more test-time compute at inference. The company initially chose to hide raw chains and instead return a model-written summary, stating that it "decided not to show" the underlying thoughts so researchers could monitor them without exposing unaligned content to end users. Commercial deployments document separate "reasoning tokens" that meter hidden thinking and a control for "reasoning effort" that tunes how much compute the model uses. These features make the models slower than ordinary chat systems while enabling stronger performance on difficult problems. == History == The research trajectory toward reasoning models combined advances in supervision, prompting, and search-style inference. Early alignment work on reinforcement learning from human feedback showed that models can be fine-tuned to follow instructions with "human feedback" and preference-based rewards. In 2022, Google Research scientists Jason Wei and Denny Zhou showed that chain-of-thought prompting "significantly improves the ability" of large models on complex reasoning tasks. Input → Step 1 → Step 2 → ⋯ → Step n ⏟ Reasoning chain → Answer {\displaystyle {\text{Input}}\rightarrow \underbrace {{\text{Step}}_{1}\rightarrow {\text{Step}}_{2}\rightarrow \cdots \rightarrow {\text{Step}}_{n}} _{\text{Reasoning chain}}\rightarrow {\text{Answer}}} A companion result demonstrated that the simple instruction "Let's think step by step" can elicit zero-shot reasoning. Follow-up work introduced self-consistency decoding, which "boosts the performance" of chain-of-thought by sampling diverse solution paths and choosing the consensus, and tool-augmented methods such as ReAct, a portmanteau of Reason and Act, that prompt models to "generate both reasoning traces" and actions. Research then generalized chain-of-thought into search over multiple candidate plans. The Tree-of-Thoughts framework from Princeton computer scientist Shunyu Yao proposes that models "perform deliberate decision making" by exploring and backtracking over a tree of intermediate thoughts. OpenAI's reported breakthrough focused on supervising reasoning processes rather than only outcomes, with Lightman et al.'s "Let's Verify Step by Step" reporting that rewarding each correct step "significantly outperforms outcome supervision" on challenging math problems and improves interpretability by aligning the chain-of-thought with human judgment. OpenAI's o1 announcement ties these strands together with a large-scale reinforcement learning algorithm that trains the model to refine its own chain of thought, and it reports that accuracy rises with more training compute and more time spent thinking at inference. Together, these developments define the core of reasoning models. They use supervision signals that evaluate the quality of intermediate steps, they exploit inference-time exploration such as consensus or tree search, and they expose controls for how much internal thinking compute to allocate. OpenAI's o1 family made this approach available at scale in September 2024 and popularized the label "reasoning model" for LLMs that deliberately think before they answer. The development of reasoning models illustrates Richard S. Sutton's "bitter lesson" that scaling compute typically outperforms methods based on human-designed insights. This principle was demonstrated by researchers at the Generative AI Research Lab (GAIR), who initially attempted to replicate o1's capabilities using sophisticated methods including tree search and reinforcement learning in late 2024. Their findings, published in the "o1 Replication Journey" series, revealed that knowledge distillation, a comparatively straightforward technique that trains a smaller model to mimic o1's outputs, produced unexpectedly strong performance. This outcome illustrated how direct scaling approaches can, at times, outperform more complex engineering solutions. === Drawbacks === Reasoning models require significantly more computational resources during inference compared to non-reasoning models. Research on the American Invitational Mathematics Examination (AIME) benchmark found that reasoning models were 10 to 74 times more expensive to operate than their non-reasoning counterparts. The extended inference time is attributed to the detailed, step-by-step reasoning outputs that these models generate, which are typically much longer than responses from standard large language models that provide direct answers without showing their reasoning process. One researcher in early 2025 argued that these models may face potential additional denial-of-service concerns with "overthinking attacks." === Releases === ==== 2024 ==== In September 2024, OpenAI released o1-preview, a large language model with enhanced reasoning capabilities. The full version, o1, was released in December 2024. OpenAI initially shared preliminary results on its successor model, o3, in December 2024, with the full o3 model becoming available in 2025. Alibaba released reasoning versions of its Qwen large language models in November 2024. In December 2024, the company introduced QvQ-72B-Preview, an experimental visual reasoning model. In December 2024, Google introduced Deep Research in Gemini, a feature designed to conduct multi-step research tasks. On December 16, 2024, researchers demonstrated that by scaling test-time compute, a relatively small Llama 3B model could outperform a much larger Llama 70B model on challenging reasoning tasks. This experiment suggested that improved inference strategies can unlock reasoning capabilities even in smaller models. ==== 2025 ==== In January 2025, DeepSeek released R1, a reasoning model that achieved performance comparable to OpenAI's o1 at significantly lower computational cost. The release demonstrated the effectiveness of Group Relative Policy Optimization (GRPO), a reinforcement learning technique used to train the model. On January 25, 2025, DeepSeek enhanced R1 with web search capabilities, allowing the model to retrieve information from the internet while performing reasoning tasks. Research during this period further validated the effectiveness of knowledge distillation for creating reasoning models. The s1-32B model achieved strong performance through budget forcing and scaling methods, reinforcing findings that simpler training approaches can be highly effective for reasoning capabilities. On February 2, 2025, OpenAI released Deep Research, a feature powered by their o3 model that enables users to conduct comprehensive research tasks. The system generates detailed reports by automatically gathering and synthesizing information from multiple web sources. OpenAI called GPT-4.5 its "last non-chain-of-thought model", and implemented with GPT-5 a router model that selects a model based on the difficulty of the task. ==== 2026 ==== In January 2026, Moonshot AI released Kimi K2.5, an open-source 1 trillion parameter MoE model with 32 billion active parameters. It uses an “Agent Swarm” system that dynamically decomposes tasks into sub-agents for reasoning and execution, enabling more scalable multi-step problem solving than a single sequential reasoning chain. == Training == Reasoning models follow the familiar large-scale pretraining used for frontier language models, then diverge in the post-training and optimization. OpenAI reports that o1 is trained with a large-
Elastic net regularization
In statistics and, in particular, in the fitting of linear or logistic regression models, the elastic net is a regularized regression method that linearly combines the L1 and L2 penalties of the lasso and ridge methods. Nevertheless, elastic net regularization is typically more accurate than both methods with regard to reconstruction. == Specification == The elastic net method overcomes the limitations of the LASSO (least absolute shrinkage and selection operator) method which uses a penalty function based on ‖ β ‖ 1 = ∑ j = 1 p | β j | . {\displaystyle \|\beta \|_{1}=\textstyle \sum _{j=1}^{p}|\beta _{j}|.} Use of this penalty function has several limitations. For example, in the "large p, small n" case (high-dimensional data with few examples), the LASSO selects at most n variables before it saturates. Also if there is a group of highly correlated variables, then the LASSO tends to select one variable from a group and ignore the others. To overcome these limitations, the elastic net adds a quadratic part ( ‖ β ‖ 2 {\displaystyle \|\beta \|^{2}} ) to the penalty, which when used alone is ridge regression (known also as Tikhonov regularization). The estimates from the elastic net method are defined by β ^ ≡ argmin β ( ‖ y − X β ‖ 2 + λ 2 ‖ β ‖ 2 + λ 1 ‖ β ‖ 1 ) . {\displaystyle {\hat {\beta }}\equiv {\underset {\beta }{\operatorname {argmin} }}(\|y-X\beta \|^{2}+\lambda _{2}\|\beta \|^{2}+\lambda _{1}\|\beta \|_{1}).} The quadratic penalty term makes the loss function strongly convex, and it therefore has a unique minimum. The elastic net method includes the LASSO and ridge regression: in other words, each of them is a special case where λ 1 = λ , λ 2 = 0 {\displaystyle \lambda _{1}=\lambda ,\lambda _{2}=0} or λ 1 = 0 , λ 2 = λ {\displaystyle \lambda _{1}=0,\lambda _{2}=\lambda } . Meanwhile, the naive version of elastic net method finds an estimator in a two-stage procedure : first for each fixed λ 2 {\displaystyle \lambda _{2}} it finds the ridge regression coefficients, and then does a LASSO type shrinkage. This kind of estimation incurs a double amount of shrinkage, which leads to increased bias and poor predictions. To improve the prediction performance, sometimes the coefficients of the naive version of elastic net is rescaled by multiplying the estimated coefficients by ( 1 + λ 2 ) {\displaystyle (1+\lambda _{2})} . Examples of where the elastic net method has been applied are: Support vector machine Metric learning Portfolio optimization Cancer prognosis == Reduction to support vector machine == It was proven in 2014 that the elastic net can be reduced to the linear support vector machine. A similar reduction was previously proven for the LASSO in 2014. The authors showed that for every instance of the elastic net, an artificial binary classification problem can be constructed such that the hyper-plane solution of a linear support vector machine (SVM) is identical to the solution β {\displaystyle \beta } (after re-scaling). The reduction immediately enables the use of highly optimized SVM solvers for elastic net problems. It also enables the use of GPU acceleration, which is often already used for large-scale SVM solvers. The reduction is a simple transformation of the original data and regularization constants X ∈ R n × p , y ∈ R n , λ 1 ≥ 0 , λ 2 ≥ 0 {\displaystyle X\in {\mathbb {R} }^{n\times p},y\in {\mathbb {R} }^{n},\lambda _{1}\geq 0,\lambda _{2}\geq 0} into new artificial data instances and a regularization constant that specify a binary classification problem and the SVM regularization constant X 2 ∈ R 2 p × n , y 2 ∈ { − 1 , 1 } 2 p , C ≥ 0. {\displaystyle X_{2}\in {\mathbb {R} }^{2p\times n},y_{2}\in \{-1,1\}^{2p},C\geq 0.} Here, y 2 {\displaystyle y_{2}} consists of binary labels − 1 , 1 {\displaystyle {-1,1}} . When 2 p > n {\displaystyle 2p>n} it is typically faster to solve the linear SVM in the primal, whereas otherwise the dual formulation is faster. Some authors have referred to the transformation as Support Vector Elastic Net (SVEN), and provided the following MATLAB pseudo-code: == Software == "Glmnet: Lasso and elastic-net regularized generalized linear models" is a software which is implemented as an R source package and as a MATLAB toolbox. This includes fast algorithms for estimation of generalized linear models with ℓ1 (the lasso), ℓ2 (ridge regression) and mixtures of the two penalties (the elastic net) using cyclical coordinate descent, computed along a regularization path. JMP Pro 11 includes elastic net regularization, using the Generalized Regression personality with Fit Model. "pensim: Simulation of high-dimensional data and parallelized repeated penalized regression" implements an alternate, parallelised "2D" tuning method of the ℓ parameters, a method claimed to result in improved prediction accuracy. scikit-learn includes linear regression and logistic regression with elastic net regularization. SVEN, a Matlab implementation of Support Vector Elastic Net. This solver reduces the Elastic Net problem to an instance of SVM binary classification and uses a Matlab SVM solver to find the solution. Because SVM is easily parallelizable, the code can be faster than Glmnet on modern hardware. SpaSM, a Matlab implementation of sparse regression, classification and principal component analysis, including elastic net regularized regression. Apache Spark provides support for Elastic Net Regression in its MLlib machine learning library. The method is available as a parameter of the more general LinearRegression class. SAS (software) The SAS procedure Glmselect and SAS Viya procedure Regselect support the use of elastic net regularization for model selection.