Hey crew, this week we’re cutting through the noise and focusing on what actually moves creators forward.

No buzzwords, just clear shifts you can use: AI that learns from your process, inboxes beating feeds, bite-sized products outselling bundles, and browser-native motion tools that let you ship in hours, not days. Think practical plays, simple setups, and one clear win per send.

Let’s dive in.

Drug discovery runs on property prediction. Before you synthesize a compound, you want to know if it'll dissolve in water, whether it'll survive long enough to reach a target, and roughly how it'll behave thermodynamically. The models doing this prediction have been, charitably, a mess of tradeoffs.

Existing molecular foundation models borrow architecture ideas from NLP and vision. They generalize okay. But they're thermodynamically inconsistent, meaning predicted properties can violate basic physical laws. Not a minor footnote. Predicting that a molecule has properties that couldn't coexist in reality is like a weather model that forecasts 100% humidity and 0% chance of rain simultaneously. Useless at best, expensive at worst.

Domain-informed models fix the physics problem but create a new one: they're built for single properties and small datasets. You get accuracy in a narrow lane and nothing else. Pick your poison.

The paper introduces MultiPUFFIN, a multimodal foundation model that tries to solve both problems at once. The architecture fuses three molecular representations: SMILES strings, molecular graphs, and 3D geometries. These get combined via gated cross-attention, which lets the model learn which representation matters most depending on context. Not every property cares equally about 3D geometry. The gate learns that.

The "domain-constrained" part is where it gets interesting. The model bakes in thermodynamic consistency directly, so predictions respect physical laws by construction rather than by luck. This sounds obvious in retrospect. It's genuinely hard to implement without sacrificing generalization.

And the multimodal fusion is doing real work here. SMILES alone is a string representation that loses geometric information. Graphs capture connectivity but not conformation. 3D geometry is expressive but expensive to compute and noisy. Combining all three, with a learned gating mechanism to weight them, is a reasonable bet on having your cake and eating it too.

The gated cross-attention fusion is the technical decision worth paying attention to. Static fusion (just concatenating representations) treats all modalities equally regardless of the task. Fine for benchmarks. Less fine when you're predicting a property that's heavily conformation-dependent versus one that's mostly determined by functional groups.

My read: the gating mechanism is what separates this from "we threw three encoders at the problem." Whether it actually learns meaningful modality weights or just collapses to one dominant representation in practice is an open question the abstract doesn't fully answer. I'd want to see the attention weights across property types before calling this solved.

So the thermodynamic consistency constraint is the other piece worth noting. Enforcing physical consistency as a hard constraint rather than a soft regularization term is the right call. Soft regularization gets traded away under pressure from the loss function. Hard constraints don't.

If this holds up under scrutiny, the downstream value is real. Physicochemical property prediction is a bottleneck at multiple stages of drug discovery: ADMET screening, formulation, and synthesis planning all depend on it. A model that's both generalizable and physically consistent would replace a pipeline that currently stitches together several specialized tools.

Materials science has the same problem at larger scale. Predicting thermal stability, solubility, and reactivity for novel materials requires exactly the kind of multi-property, thermodynamically consistent prediction that single-property models can't provide.

Not a solved problem yet. But directionally correct.

I've watched a lot of molecular ML papers promise multimodal fusion and deliver marginal benchmark improvements. MultiPUFFIN's core insight, that you need to enforce physical consistency architecturally rather than hope it emerges from training data, is the right framing. Most foundation models in this space treat chemistry like text prediction. Chemistry has laws. Ignoring them produces models that are confidently wrong in ways that waste expensive lab time.

But the paper is an arXiv preprint (submitted March 2025), and the summary cuts off mid-sentence. I can't evaluate the actual benchmark numbers or dataset scale, which are the things that would tell me whether this is a genuine step forward or a well-motivated architecture with underwhelming empirics.

The approach deserves serious attention. The proof is still pending.

--- The AI finds the signal. We decide what it means.

Notion has quietly become the tool I'd recommend to anyone drowning in scattered docs, half-finished wikis, and project trackers living in four different apps. The AI features aren't just bolted on. They actually help: summarizing meeting notes, drafting content briefs, querying your own database in plain English.

Worth trying if you're a founder or small team who needs a single place for docs, projects, and knowledge without paying for five separate tools.

The learning curve is real. Give it a week before judging it.

But once it clicks, you won't want to go back. Start free at notion.so.

This is an affiliate link. We earn a commission if you subscribe. We only feature tools we'd use ourselves.

Some links in Tool Radar are affiliate links. We earn a small commission at no extra cost to you.

Bigger is smarter. That's the assumption baked into how most people talk about AI progress. A 70B model obviously beats a 7B model. A trillion-parameter model must be god-tier. Just scale it up.

Wrong.

People believe this because the early scaling laws made it look true. From 2019 to 2022, bigger models pretty reliably won benchmarks. The correlation was real enough that "how many parameters?" became shorthand for "how good is it?" Reasonable inference from the data at the time.

But Mistral 7B outperformed GPT-3.5 on several benchmarks in late 2023. A model roughly 25x smaller by parameter count. And it wasn't a fluke. Phi-2, at 2.7 billion parameters, beat models 10x its size on reasoning tasks. The reason: training data quality and architecture choices matter more than raw size past a certain threshold. Saying a bigger model is always smarter is like saying a bigger hard drive makes your laptop faster. Different thing entirely.

So what actually predicts model quality? Compute efficiency, data curation, fine-tuning approach, and increasingly, inference-time reasoning (how much the model "thinks" before answering). Parameters are one input. Not the scoreboard.

And the industry knows this. The race right now isn't to a trillion parameters. It's to smaller models that run locally, cheaply, and fast. Qualitatively better. Quantitatively smaller.

The one-sentence version: Parameter count is a proxy metric that stopped being reliable the moment people started treating it as the actual metric.

Claude Opus 4.6 →
Anthropic's sharpest model yet, built specifically for agentic workflows.

Qwen3.5-9B →
A 9B hybrid attention model you can run locally. Worth watching.

Google Nano Banana 2 →
Google's new image model ships today, replacing its predecessors immediately.

Gemini 3.1 Pro →
Google claims better complex reasoning. Ars Technica has the details.

Sonnet 4.6 →
Anthropic's mid-tier model gets a meaningful capability bump.

Gemini 3.1 Pro on Product Hunt →
Community reactions to Google's latest, straight from early adopters.

Tools gaining traction this week based on our source data. Some affiliate links.

This is an affiliate link. We earn a commission if you subscribe. We only feature tools we'd use ourselves.