News

Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline

NVIDIA and IIT Madras' AI4Bharat developed Nemotron-pretraining-Indic-v1, an approximately 2.8 trillion-token corpus. It includes native-script data for 22 Indian languages and romanized text for twelve. Together, those cover 34 language-script configurations for pretraining. The corpus is described as releasing soon.

Each generated record keeps a link to its source

The open generation pipeline uses GPU acceleration. It combines document-level translation with quality assessment and targeted repair. Its checks are designed to preserve equations, tables, and code. Every corpus record traces back to a curated source document. Each output is checked against its input. Continued pretraining for Nemotron 3.5 Super VL added 10 billion Indic tokens from a twelve-language subset of the corpus and natural Indic text. (Hugging Face)

thinkidiot take: Every corpus record traces back to a curated source document, and every output is checked against its input. I would use that traceability to inspect how translations handle equations, tables, and code before choosing training data. The targeted repair stage also gives quality assessment a direct role in generation. For my own training work, that inspection path matters more than the 2.8 trillion-token headline.

How Meta Built Forum With AI-Native Engineering

Meta built Forum, a standalone Groups app, from scratch with a small team using an AI agent. The app is for people who want to go deeper into conversations in their Facebook groups. Its development process gives the agent responsibility for planning, implementation, building, testing, and correction.

A finished task needs evidence and an engineer's approval

Engineers set goals, supervise the agent, and approve results. Some stages still require an engineer in each iteration. The team measures success through production-ready changes that build, pass tests, and survive review. Tasks count as complete only after automated tests and visual checks pass, and every shipped outcome requires an engineer's sign-off. Build status, unit tests, and matching against UI specifications are integrated into the agent loop. Changes carry build results, test outcomes, and recorded verification sessions. (Meta Engineering)

thinkidiot take: Every shipped Forum outcome requires an engineer's sign-off, even with the agent handling much of the development cycle. I would borrow the practice of attaching build results, test outcomes, and recorded verification sessions to each change. That gives the reviewer concrete evidence to assess alongside the implementation. I judge this workflow by its definition of finished work, and Meta has chosen the right one.

Accelerating the search for antimicrobial peptides with Google Antigravity

Google DeepMind presented a walkthrough using Google Antigravity for Science to search for novel antimicrobial peptides. The work targets Streptococcus pseudopneumoniae. Its agentic workbench supports discovering biological targets, evaluating their structures, and prioritizing them for further work.

The selected candidate is now undergoing wet-lab validation

ProtNLM predicts biological function from raw amino acid sequences. It has added functional labels to millions of previously unnamed UniProt entries. Its annotations cover protein names, function comments, subcellular locations, GO terms, and keywords. In the walkthrough, the researcher steers strategy while the agent handles work across databases, structural analysis, and literature synthesis. The identified peptide candidate has predicted activity against pathogens involved in respiratory tract infections. That candidate is undergoing wet-lab validation. (Google DeepMind)

thinkidiot take: The identified peptide has predicted activity and is undergoing wet-lab validation. I would use this workbench to assemble structural analysis and literature evidence for deciding which targets deserve testing. The researcher still steers strategy while the agent coordinates that supporting work. My judgement is that candidate selection is the useful achievement here; the biological result belongs to the lab.

Trust in Depth: Giving agents more autonomy without losing control

AWS described Trust in Depth, a layered approach to giving agents autonomy while retaining control. It recommends autonomous action when actions can be verified and reserves human judgment for decisions that require it. The framework organizes controls around four lifecycle questions, giving teams a structure for deciding where an agent acts independently.

A proof only covers the system and assumptions it describes

Scope comes before execution, and Know applies at deployment. Bound applies during execution, while Observe follows each run. Automated reasoning can prove that a precise rule holds on every path covered by a system description and its assumptions. It can also identify a case where the rule fails. AWS says it has used automated reasoning in production for more than a decade, including IAM at more than 2.1 billion requests per second. Its other examples include AWS Nitro Hypervisor, Amazon S3 Block Public Access, and VPC Reachability Analyzer. (AWS Machine Learning)

thinkidiot take: Automated reasoning proves rules within a specified system and its assumptions. I would write down that scope before using a proof to justify autonomous action. Decisions outside that scope still require their own basis for trust. For me, explicit assumptions are the most valuable part of this approach because they define exactly what the verification establishes.

When it came to Cal AI’s viral success, a small army of influencers paid off

Cal AI, an AI-powered calorie tracking app founded in 2024, generated over $30 million in revenue in two years. MyFitnessPal acquired it in March for an undisclosed amount. Founder Zach Yadegari said an early influencer post brought approximately 1,000 new users in two days.

Yadegari's next venture puts a personal assistant across communication channels

Cal AI negotiated influencer payments against reach targets. It sought influencers with tightly connected communities. Yadegari launched AI agent startup Persona earlier in October 2026. The startup had already raised $10 million. Persona plans a personal AI assistant accessible through iMessage, an app, email, and phone calls. Its plans also include a hardware ecosystem around the app. (TechCrunch)

thinkidiot take: An early influencer post brought Cal AI approximately 1,000 new users in two days, according to Yadegari. I would use that concrete acquisition result as the starting point for assessing the influencer strategy. Negotiating payments against reach targets gives the spending an explicit target, but reach and new users are different measures. The user count is the more persuasive evidence to me.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-10-16.

OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding

OmniCompass tests how well AI systems can follow a video. It checks whether they can locate something on screen, identify when it appears, and explain what happens. Some tests require pictures alone, while others require sound too. The aim is to reveal weaknesses that narrower tests miss.

• Problem: Existing tests examine only some parts of video understanding. That makes it hard to tell whether a model struggles with timing, location, events, or the connection between sound and pictures.
• New idea: OmniCompass splits its tests into a visual track, where answers need evidence from pictures, and an audio-visual track, where answers need both sound and pictures. Each track tests temporal grounding, meaning finding when a target appears, and spatial grounding, meaning finding where it appears on screen. It also asks questions about what happens. Each target is described twice, once with text alone and once with text plus an image, to check whether the model gives consistent answers.
• Simple example: Think of asking someone to find a particular moment in a home video. They need to say when it happens, point to the right part of the screen, and explain the event. You could describe the target in words, then add a picture and see whether their answer changes.
• Evidence: The visual track contains 820 videos and 2,460 instances. The audio-visual track uses a subset of 240 videos and 720 instances. Tests of 50 models found that open-source models trailed proprietary models, especially on tasks needing sound and pictures. Even leading models struggled to locate targets precisely in time and space, and answers were not always consistent between text-only and text-plus-image prompts.
• Limitation: The abstract reports broad performance gaps but gives no scores or gap sizes, so it does not show how large those differences are.
• Why it matters: A model's ability to describe a video does not establish that it can reliably locate the evidence behind its answer.
• Paper: OmniCompass: A Holistic Benchmark for Omni-Modal Video

Speculative Forcing for Efficient Autoregressive Video Generation

Speculative Forcing aims to make AI video generation faster by stopping work on parts that are already ready. Video generators refine their output in stages, but some parts settle sooner than others. This method tries to recognize those parts and leave them alone. The remaining parts keep receiving attention.

• Problem: The video generators studied here give every token, a small unit of generated content, the same number of refinement steps. That wastes work when a token already closely resembles the result it would reach after all the steps.
• New idea: Speculative Forcing treats unfinished estimates of generated content as drafts that can be accepted early. A verifier, a check that decides whether a draft is ready, assesses each token separately. The paper tests two verifiers: one uses similarity to clean content and stability between steps without extra training, while the other learns to estimate how far a token remains from its fully refined result. Accepted tokens stop changing, but their stored information stays available to guide the remaining work.
• Simple example: Think of editing a paragraph where some sentences are already finished and others still need work. You stop revising the finished sentences but keep reading them for context as you fix the rest.
• Evidence: On LongLive, the method matched the four-step generator's quality using about 1.7 denoising steps per token, meaning rounds of refinement that remove noise. It ran over 1.5 times faster, with reported speeds of 41.7 and 37.8 frames per second. On Lingbot-World 2.0, it maintained quality with about 2.5 steps per token and ran 1.35 times faster at 24.5 frames per second.
• Limitation: The abstract says quality was preserved but does not name the quality measures or report how often a token was accepted too early.
• Why it matters: Stopping refinement when a part is ready can make video generation faster while preserving the reported output quality.
• Paper: Speculative Forcing for Efficient Autoregressive Video

Orbia: Do Generated Videos Form a Consistent 3D World?

Orbia tests whether an AI-generated video holds together as a three-dimensional place. A scene can look convincing in one view without making sense when the viewer moves around. The tests explore places and return to them to check what stays the same. They examine both the scene supplied at the start and the surroundings the model invents.

• Problem: Realistic pictures do not establish that a generated world has stable shapes, positions, or surroundings. Existing tests check selected geometric properties or ask AI models to judge consistency, leaving gaps in how thoroughly they examine the underlying space.
• New idea: Orbia organizes its checks around preserving the starting scene, retaining newly invented content, and keeping different views compatible with one three-dimensional place. It checks the starting scene against reference images paired with depth measurements, which record distances, and marked reference points. It returns to areas the model invented and compares them with their first appearance. It also checks whether distances, matching visual details, and the scale of regions agree across views, including during minute-long explorations and returns.
• Simple example: Imagine exploring a generated room, turning a corner, and then walking back. The room should retain its layout, and anything first revealed around the corner should still be there when you return. Orbia uses this kind of exploration and return to check whether the place stays consistent.
• Evidence: Orbia contains 720 instances covering indoor and outdoor environments, with six procedures for exploration and return visits. The evaluation covers 27 models guided by text, actions, or camera controls. It also includes minute-long exploration and return tests, but the abstract reports no model scores or rankings.
• Limitation: The abstract describes the tests and their coverage but gives no performance results showing which models maintain a consistent world or where they fail.
• Why it matters: Generated worlds used for exploration and simulation need to remain coherent as the viewer moves and returns.
• Paper: Orbia: Do Generated Videos Form a Consistent 3D World?

Trending AI Repositories

Ranking source: GitHub Trending.

lightseekorg/tokenspeed

TokenSpeed focuses on running models for production agents. Its stated goal pairs TensorRT-LLM-level performance with vLLM-level usability, making that balance the main point to examine.

• What it is: TokenSpeed runs LLM inference for agentic workloads. Its listed components include a scheduler with a C++ control plane.
• What it does: TokenSpeed is a speed-of-light LLM inference engine.
• Who it helps: It targets developers serving models for agentic workloads. They can evaluate it against the project's stated performance and usability goals.
• Limitation: The supplied excerpt gives no benchmarks to support the performance claim.
• Repository: lightseekorg/tokenspeed

xerj-org/xerj

XERJ is a Rust project focused on data retrieval for AI agents. Its stated reduction in token use makes retrieval cost the main point to examine.

• What it is: XERJ indexes code, documents, logs, and PDFs for agent search and retrieval. The project claims a 40x reduction in token use and compatibility with Elasticsearch.
• What it does: XERJ is the new way for AI to search data. Its autoindex capability activates agents to know your data without the token waste of grep and sed. One command indexes code, docs, logs and PDFs for search, RAG, security audits and agent memory, using 40x fewer tokens than grep. Elasticsearch compatible, so existing clients just work.
• Who it helps: It serves developers building agents that need access to their own data. They can use the indexed material to support retrieval and agent memory.
• Limitation: The supplied material gives no benchmark setup to support the token-saving claim.
• Repository: xerj-org/xerj