NeurIPS 2023, Part 4 of 4: Making AI Easier to Build and Understand

Keeping open source collaborative under economic pressure

Economic pressure can strain open-source collaboration. Nomic closes its conversation in Latent Space’s NeurIPS 2023 recording with a concrete commitment: keep GPT-4 All free for its community, improve it in response to feedback, and contribute to other high-impact open-source projects. Preserving collaboration as those pressures increase is the stated goal.

Contributing also offers a path into the team. Nomic encourages people to make maps, join its Discord discussions, and do open-source work; its representative says substantial contributors accounted for many hires in the recent hiring round. The closing contact points are Nomic_AI on Twitter and Nomic.ai.

Lightning: Control the training loop, reduce expensive mistakes

The next interview introduces Luca, Lightning.ai’s CTO at the time, at the launch of Lightning Studio, a cloud development environment. He describes PyTorch Lightning as a training framework adapting to generative AI, citing Stable Diffusion’s training and NVIDIA NeMo as examples of its use. He reports roughly 5.5 million downloads per month and 80 million in total at the time of the interview.

Large models raise the cost of mistakes. Luca says releases 2.0 and 2.1 focused on memory pressure and distributed training—training spread across multiple machines—at larger model sizes. Running out of memory is an obvious failure; incorrectly combining accuracy measurements across machines is subtler, because the resulting metric can mislead the developer. His rationale for frameworks is reducing opportunities for mistakes that waste compute.

Lightning Fabric changes which parts of training the developer controls: - Lightning Trainer supplies the loop structure. Bespoke behavior remains possible, but unconventional loops can be harder to express. - Fabric lets developers write their own loops. It exposes the Trainer’s components while retaining abstractions for strategies and precision plugins. Luca highlights reinforcement-learning loops and custom placement of collective communication calls as reasons to choose it.

Fabric also powers the open-source language-model repositories Lit Llama and LitGPT. Luca describes adapting NanoGPT into a minimal, single-file Llama implementation so developers could inspect and modify model internals without navigating layers of framework code. By the interview, LitGPT supported multiple models and appeared in the large language model (LLM) efficiency challenge starter kit.

The development principle is make code easy to modify before optimizing it. Accessible implementations invite contributions, and practical use can expose improvements worth incorporating into the shared frameworks. The discussion cites TinyLlama and a report that 360 AI had used Lit Llama to pretrain a 7-billion-parameter model, with lessons feeding back into Fabric and PyTorch Lightning.

TinyLlama provides the concrete feedback example: a 1.1-billion-parameter model with a 3-trillion-token training plan, which the speakers believed was still underway. The team reports bringing optimizations back into its code, tentatively recalling chunked cross-entropy and some kernels. Its dataset implementation, which divides data into chunks for fast streaming, also worked for TinyLlama.

The team also describes native bitsandbytes support in Fabric and PyTorch Lightning as extending quantization support across model types, beyond architectures selected by the original authors. This is a claim about flexibility; no performance comparison accompanies it.

Lightning Studio: Change compute without rebuilding the workspace

Lightning Studio extends the same concern about mistakes to the development environment. Users open VS Code in a browser or connect from local VS Code and work on a cloud machine, without first having to build a Docker container.

The central promise is workspace persistence when compute changes. Installed dependencies reportedly carry over when switching instance types. Duplicating a studio preserves its environment, data, code, and saved checkpoints. The underlying persistence mechanism is not explained; the interviewer’s suggestion of memory snapshotting remains a question.

The design test is concrete: a developer who needs four GPUs should be able to acquire them and continue working. Provisioning machines, rebuilding environments, and attaching drives interrupt that flow and introduce more opportunities for error. The priority is developer velocity through fewer infrastructure chores.

The intended workloads range from small experiments with persistent data across restarts to large language model training. The team says a job can run on 100 GPUs provided the machines are available. The platform can manage reservations or orchestrate workloads inside a customer’s cloud account, where the customer negotiates provider quotas.

At the NeurIPS 2023 launch, infrastructure support had three distinct stages: - AWS was the current focus, with the provider reportedly supporting the launch. - Other clouds were planned. The team had spoken with Oracle and other providers, but did not announce available integrations. - On-premises orchestration was forthcoming, including support for clusters using the Slurm scheduler. Work with some companies was underway, but public support had not shipped.

An efficiency challenge needs reliable evaluation

Lightning’s motivation for supporting the LLM efficiency challenge is that smaller models can give people more control and help them learn to extract useful results from AI. The interviewee argues that efficiency and shorter iteration cycles could help open source match or exceed closed-source capabilities—a hoped-for outcome.

Its concrete contribution was a Discord-connected public leaderboard, with a simple submission sequence: 1. A participant sent a submission by direct message to a bot. 2. An agent running on several machines queued and launched the job. 3. The service collected evaluation results and produced a ranking.

Operating the service was harder than that interface suggested. The first month was rough: the infrastructure was new, built in spare time, and improved with community feedback. More than 1,000 submissions were reportedly fully evaluated; many others failed to complete because of submission or user-code problems. The count measures completed evaluations, not strong model performance.

Other organizers subsequently reevaluated submissions and reran leading entries from scratch, reusing some leaderboard code. The interviewee had been occupied with the Studio launch and had not participated in that work, so offered no analysis of the winning methods.

The proposed next step was public studios containing evaluation frameworks, where community members could spend free credits evaluating their own models. An energetic community discussion analyzing Mixtral after its release helped motivate the plan.

The Lightning interview closes with advice for finding ideas: visit less crowded posters and talk candidly with their authors, including those working outside your immediate focus. The interviewer recounts an unnamed presenter privately expressing doubt about their own paper after the field had moved on since submission. The anecdote concerns a researcher’s changed view, rather than a finding about the paper’s validity.

Jay Alammar: What if token five is already wrong?

The final interview turns to Jay Alammar and education for practitioners. He welcomes AI engineering as a useful category of work, clarifying an audience’s needs as machine learning moves from research into industry. At NeurIPS 2023, he sees growing interest in LLMs and highlights diffusion, including diffusion for text, as a research direction to watch.

Asked why text needs diffusion when users may want streaming output, Alammar points to early commitment in autoregressive generation: text is generated successively, so a mistake at token five remains in the sequence as generation continues. He is interested in whether avoiding that commitment could enable different applications; he does not present a comparative result.

The interviewer proposes Tree of Thought as a remedy, and Alammar acknowledges it as one avenue. A participant then recounts hearing from someone at Midjourney about a storytelling group exploring text diffusion. Its motivation was coherence across an entire story, analogous to composing an image as a whole. This is a secondhand account of exploration.

The discussion also anticipates reasoning research influencing agents quickly. Chain of thought and tree of thought are named as promising directions, while ReAct’s adoption in LangChain supplies a concrete example of research moving into agent tooling. Agents are still described as early in their development.

Finding researchers and investigating model knowledge

Navigating NeurIPS is itself an information problem. Paper visualizations help attendees orient themselves, while oral sessions and personal encounters expose work they might otherwise miss. One participant discovered DataComp through an oral presentation, describing it loosely as a new benchmark in the spirit of ImageNet. Another recounts meeting a researcher and checking their Google Scholar profile, only to realize their papers had been familiar for years even though their name was not.

David Bau’s work on interpretability and editing language-model knowledge provides a technical example. Two questions stand out: - Can a stored association be changed? The example edits a model to associate the Eiffel Tower with Rome instead of Paris. Repeated edits can degrade a model; 100 edits is illustrative, not a measured threshold. Subsequent work is mentioned as addressing larger numbers of edits. - Where does a token prediction emerge? LogitLens is described as a way to investigate which layers begin suggesting a particular token.

Publishing researcher interviews through Cohere’s YouTube channel and social accounts is described as learning in public. Work with customers of Cohere’s embeddings, reranking, and generation models provides another perspective: industry applications reveal what people are trying to accomplish and where they need help using models effectively.

Teach Transformer improvements that remain useful

Asked about state-space models and what might follow Transformers, the respondent has not yet formed a view. An overlooked poster could become influential months later. The participants recall the original Transformer as a poster-only paper initially associated with machine translation, and invoke BERT and GPT to illustrate its broader relevance. The interviewer wants a survey tracing the evolution from Attention Is All You Need toward the modern decoder-only paradigm.

The planned update to The Illustrated Transformer uses sustained usefulness as its selection criterion. The discussion names RoPE and ALiBi positional encodings, FlashAttention, localized attention, and multiquery or grouped-query variants as material for the update. The speaker regards the vanilla Transformer as having outlasted many proposed enhancements: an improvement on one problem is less persuasive than repeated adoption over time.

That criterion has a concrete boundary: sliding-window attention was not yet included as an established improvement. The speaker cited its use in Mistral but wanted to see adoption in more work. This was a judgment at the time of the 2023 interview. The updated explanation was planned as a chapter in the O’Reilly book Hands On Large Language Models, then being co-written.

Discover papers, meet people, keep learning

The closing conference advice offers complementary ways to find relevant research: - Browse topic clusters. A clustered paper visualization credited to Hendrik Strobelt and Ben Hoover is recommended for orientation and attendance planning. - Search by meaning. Semantic search can find papers about agents without requiring exact keyword matches. - Ask for relevant papers. A Cohere demo over NeurIPS papers accepts interests such as LLM efficiency and returns paper recommendations.

The interviewer wants similar discovery tools for meetups, including gatherings for Singaporeans in tech and running groups. Conferences also serve a social purpose beyond technical sessions. Alammar closes by pointing listeners toward his Twitter and YouTube presence and LLM University, at LLM.university.

The outro ends both the NeurIPS coverage and Latent Space’s 2023 podcast season. Listeners are invited to complete a survey about what they liked and wanted to hear in 2024. Episodes for the new year had already been recorded, ahead of the podcast’s first anniversary.