“A lot of people who are ‘social media’ stars aren’t considered to be ‘real’ stars, and people underestimate the amount of work it takes to edit and upload a video every single day and document your life like that.”
— SHAWN MENDES
@boy_in_prime The Ikorodu to Ojota/Ikeja struggle is incredibly real, between the unpredictable price hikes and the exhaustion, the Lagos "commute tax" is way too high. 😩
It’s one thing to get a job and another thing to commute daily in Lagos
Please if you ply Ikorodu to Ojota/Ikeja with a car, kindly help my destiny. The transport cost is huge.
I would appreciate your repost too
LAGOS NIGERIA!
I’m announcing my first ever headline shows here, and we have made sure this homecoming will be special. Sign up for early access here! https://t.co/6RTJy6ubeT
Many people from computing or software backgrounds assume biology works like computer code — something clean, modular, and easily programmable. But biology is fundamentally different: it is highly complex, nonlinear, dynamic, and multidimensional, shaped by evolution, stochasticity, context, and emergent interactions across scales. Living systems don’t behave like deterministic software; they adapt, compensate, and evolve in ways that often defy simple engineering assumptions.
A multimodal LLM that speaks both atomic structure and natural language
Materials discovery needs two things at once: quantitative reasoning about atomic structures (energies, bandgaps, stability) and linguistic interaction (synthesis, candidate comparisons, literature). Universal interatomic potentials like CHGNet or MACE excel at the former but cannot read a prompt. LLMs handle language but predict numerical properties poorly, because text serializations of crystals (CIFs, SMILES) lose 3D geometry. The challenge is multimodal: one model that ingests an atomic graph AND a textual query, and answers in natural language.
Yingheng Tang and coauthors propose MatterChat, a multimodal architecture linking a pretrained graph encoder (CHGNet or MACE) to a frozen Mistral 7B LLM through a small trainable bridge inspired by BLIP-2. It uses 32 learnable query vectors and alternating self and cross-attention to turn atom-level embeddings into a representation the LLM consumes alongside the user's text. Encoder and LLM stay frozen; only the bridge trains, inheriting structural knowledge from one side and linguistic reasoning from the other.
On 142,899 inorganic crystals from the Materials Project across 12 tasks (formula, space group, stability, bandgap, formation energy, magnetism), MatterChat beats Gemini, GPT-4o and DeepSeek on formation energy prediction, and outperforms SchNet, CHGNet and MACE on classification accuracy and regression RMSE. A multimodal retrieval step over bridge embeddings cuts RMSE by 12%. It also produces literature-aligned synthesis protocols for GaN and yttrium iron garnet, and its embeddings cluster compounds by structural similarity and formation energy without explicit supervision.
For energy storage, semiconductors or catalysis, the takeaway is that you can keep your favourite atomistic model and your favourite LLM, train a small bridge, and get one interface for property screening and natural-language reasoning about synthesis. A low-friction path to AI assistants in discovery pipelines without giving up the physics-based models teams already trust.
Paper: Tang et al., Nature Machine Intelligence (2026) — CC BY-NC-ND 4.0 | https://t.co/ekfQVeoIZ7