Observed Signal · May 8, 2024 · Technical Release · Source: Trending Topics · Impact: 2/5 · Sentiment: Positive
NXAI unveils xLSTM, a new rival to GPT
NXAI, an AI startup co-founded by Sepp Hochreiter, has published a scientific paper introducing xLSTM, an extended LSTM architecture that aims to compete with transformer-based models like GPT-3. The paper, developed with researchers from the Institute for Machine Learning at JKU Linz, describes exponential gating and modified memory structures that improve performance and scaling compared to current transformers and state space models. xLSTM was trained on 15 billion and 300 billion tokens (from SlimPajama) and tested against models like Llama, Mamba, and RWKV-4. The authors claim xLSTM performs favorably in language modeling and has potential for reinforcement learning, time series forecasting, and physical system modeling. NXAI intends to deliver industrial applications that other large language models cannot provide.
New LLM architecture from an AI startup could influence future AI applications in adtech, but is not an immediate industry shift.
Track Real-Time Large Language Models Signals & Market Shifts
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- NXAI, co-founded by Sepp Hochreiter, published a paper on xLSTM, an extended LSTM architecture.
- xLSTM uses exponential gating and modified memory structures to compete with transformers.
- The model was trained on 15B and 300B tokens and compared with Llama, Mamba, and RWKV-4.
- Researchers claim xLSTM performs comparably to transformers and state space models in language modeling.
- NXAI aims to provide industry-specific solutions beyond what current LLMs offer.
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
GPT-5.6 Proves 30-Year Convex Optimization Lower Bound
On July 17, 2026, a Hacker News-linked Reddit thread reported that GPT-5.6 Sol, guided by a carefully constructed prompt, produced a complexity-theoretic proof that convex optimization over a standard bounded Lipschitz function class requires Omega(d^2) function evaluations, closing a 30-year theoretical gap. The computation took approximately 148 minutes and was human-verified by a domain expert. The article places this result alongside a recent claim that GPT-5.6 Sol Ultra (using 64 parallel subagents) produced a proof of the Cycle Double Cover conjecture, and discusses implications for attribution, research agendas, verification infrastructure (e.g., Lean formalization), and model selection for sustained formal reasoning versus architectural judgment.
Meta reportedly shifting to closed-source LLMs with 'Avocado' by 2026
Meta Platforms is reportedly planning to abandon its open-source approach for its upcoming large language model, codenamed 'Avocado', expected in 2026. The shift, driven by competition with OpenAI and Google, would mark a major strategic change. Meta has invested $14.3 billion in Scale AI and appointed its founder, Alexandr Wang, as Chief AI Officer. Long-time Chief AI Scientist Yann LeCun is leaving the company, and hundreds of employees have been laid off from the FAIR unit. The new model is expected to generate revenue, likely through a proprietary API, while older models may remain open-source under a freemium model.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.