Answer.ai

Answer.AI is an AI research and development company founded by Jeremy Howard and Eric Ries to turn research breakthroughs into practical products. Its work includes adapting large pretrained models on consumer GPUs, the ModernBERT encoder collaboration, and software tools such as FastHTML and Solveit. At its December 2023 launch, the company prioritized useful applications of existing models rather than building AGI.[2][3][6][5][9]

Howard and Ries launched Answer.AI on December 12, 2023, with Howard as founding CEO and Ries as founding director. The launch announced a USD10 million investment from Decibel VC supporting the company’s creation and a planned fully remote team of technical generalists. Ries brought experience from the Lean Startup movement and the Long-Term Stock Exchange.[2]

The accessibility mission preceded Answer.AI: Howard had co-founded fast.ai with Rachel Thomas, working on education, transfer learning, and fine-tuning. In his launch account, he described Answer.AI as sharing fast.ai’s mission while pursuing a different method—commercial products and services whose proceeds could support further development and lower costs. This earlier work supplied technical and educational context for the new company.[2]

Answer.AI’s founding approach linked research to application development. Ries summarized it as “development should inform research and vice-versa.” The company proposed solving smaller practical problems through useful intermediate steps, including better fine-tuning of smaller models and reducing barriers to AI use.[2]

On March 6, 2024, Answer.AI released its first project, FSDP QLoRA. The team reported fine-tuning a 70-billion-parameter model on two 24GB RTX 3090 GPUs, bringing large-model adaptation within reach of desktop hardware. The project was a collaboration with Tim Dettmers and Hugging Face’s Titus von Koeller and Sourab Mangrulkar. Its script was explicitly an alpha/preview release.[3]

QLoRA keeps a quantized pretrained base model frozen and trains small, unquantized low-rank adapters. Fully Sharded Data Parallel (FSDP) distributes model parameters across GPUs, gathering the required shards for each layer’s computation. A 70-billion-parameter model needs roughly 140GB for 16-bit weights or 35GB for four-bit weights, before additional training memory. Combining quantization with sharding addressed both the size of the weights and the memory limit of each consumer card; the demonstrated work adapted pretrained models.[3]

Making the techniques work together required changes across their supporting libraries. Johno Whitaker supplied a standalone test script. Benjamin Warner and von Koeller independently identified selectable storage types as a way to reconcile quantized weights with FSDP’s floating-point requirements; Warner and Kerem Turgutlu packaged the integration for bitsandbytes. Howard built a minimal training script, while Warner developed layerwise loading to avoid placing the entire model on one GPU. Warner also fixed repeated quantization when offloaded weights returned from CPU memory—a bug that corrupted the pretrained weights.[3]

The next release, FSDP QDoRA, was dated April 22, 2024 and led by Turgutlu. Answer.AI supplied quantized, FSDP-compatible implementations of other researchers’ DoRA and Llama-Pro methods. DoRA separates weight-update magnitude and direction; Answer.AI’s implementation froze and quantized pretrained layers while training QDoRA layers representing approximately two percent of model parameters. Its quantized Llama-Pro implementation instead added trainable decoder blocks between frozen layers.[4]

In preliminary Llama-2-7b experiments on Orca-Math, the team reported QDoRA approaching or exceeding full fine-tuning. With 10,000 training examples, exact-match scores on 500 held-out examples were 17.6% for QDoRA, 9.8% for QLoRA, and 18.2% for full fine-tuning; with 100,000 training examples, they were 31.2%, 11.8%, and 26.0%. These were task-specific results with limited hyperparameter optimization. The team also described its inference integration as experimental and still needing optimized kernels.[4]

In an interview published August 16, 2024, recorded approximately six weeks earlier, Howard described Answer.AI as a public benefit corporation whose stated benefit was “to maximize the benefit to society through using AI.” He said its second financing round was an angel round conducted through a long-term special-purpose vehicle, with investors required to vote according to long-term-value principles.[1]

Howard described an organizational experiment with no managers or required meetings, optional meetings across major time zones, and recruitment based on demonstrated work rather than credentials. His recruiting goal was 12 people, and creating thousands of commercially successful products was an ambition. As an example of autonomy, he recalled Benjamin Clavié assembling an outside collaboration to update BERT without seeking his permission, followed by Benjamin Warner contributing a readily modifiable transformer implementation.[1][6]

Howard started FastHTML at Answer.AI to simplify application development, including for students struggling to turn AI projects into working products. The August 3, 2024 announcement described a Python framework that could begin as a single file and grow through the same underlying web technologies. It also acknowledged that the framework was young and some capabilities were not yet ready out of the box.[5]

FastHTML generates HTML through Python components, using Starlette and Uvicorn for its ASGI foundation and HTMX for browser interactions. The server can return an HTML partial that updates part of an existing page, avoiding the need for a separate JavaScript frontend. SQLite support comes through Fastlite, while other storage systems remain usable. FastAPI was a design inspiration, and its creator Sebastián Ramírez advised Howard on its design.[5]

On December 19, 2024, Answer.AI, LightOn, and collaborators released ModernBERT, an encoder-only masked language model with an 8,192-token context. It brought newer transformer engineering to BERT-style models used for classification, retrieval, and other language-processing tasks. The base model predicts masked words; downstream uses require task-specific adaptation. LightOn sponsored the project’s compute on Orange Business Cloud Avenue.[6][7]ⁿ¹

ModernBERT combines rotary positional embeddings and GeGLU layers with alternating attention: every third layer attends to the full input, while the others attend locally to the nearest 128 tokens. Unpadding and sequence packing reduce work spent on filler tokens, with masking keeping samples separate. Its dimensions were selected with common inference GPUs in mind. The release reported training on two trillion tokens of English web documents, code, and scientific articles: 1.7 trillion at short context, 250 billion for long-context adaptation, and 50 billion for final annealing.[6]

In January 2025, Hamel Husain, Isaac Flath, and Johno Whitaker published their month-long evaluation of Devin. They reported 14 failures, three successes, and three inconclusive outcomes across 20 tasks. Early successes included moving Notion data into Google Sheets with only a few minutes of human interaction during an hour-long process, and Whitaker creating a planet tracker through his phone. A Braintrust integration produced overly complex code, which they abandoned in favor of incremental work in Cursor. The authors chose to continue with tools that let developers direct the process while receiving AI assistance.[8]

Howard had previewed an experimental “AI magic” system for “dialogue engineering” in the 2024 interview. A later Solveit announcement, dated October 2, 2025, described a platform alongside fast.ai’s How to Solve it With Code course. Howard and Whitaker had taught a trial course in late 2024. Whitaker reported 1,000 preview users over the preceding year and 1,000 trial-course participants; these figures describe participation in potentially overlapping groups.[1][9]

Solveit supports short, executable steps with notes, plans, and outputs visible to both the human and AI. Users can edit dialogs, remove dead ends, hide messages from the AI, or pin important context. Python functions can become tools available to the AI. Generated code appears in fenced blocks and is neither added to the program nor executed until the user chooses, keeping review and understanding within the workflow.[9]

Whitaker reported that Answer.AI used Solveit for software development, system administration, GUI iteration, legal contract drafting, and production application hosting. The team also built and operated Solveit’s own horizontally scalable, multi-server hosting platform entirely through Solveit.[9]