Transcription
[Music] This week in AI honestly felt like a tech fever dream. We got a tiny model from Samsung that somehow beat giants like Gemini and Deep Seek at reasoning. Microsoft dropped a neural brain for quantum chemistry. Anthropic built an AI that literally audits other AIs for bad behavior. Liquid AI launched an ondevice model that could run on your phone. And Meta just re-engineered how multimodal search even works. It's like every lab decided let's confuse the internet at the same time.
All right, let's start with the weird one cuz this story sounds fake. Samsung's research lab in Montreal built a model called tiny recursive model or TRM. It has 7 million parameters which in AI terms is like bringing a water pistol to a tank fight. And yet it beat models with billions of parameters at actual reasoning. On the ARC AGI1 test, this baby model scored 44.6 to 45% and on ARC AGI 2, it scored around 8% when given two tries. For comparison, DeepSeek R1 scored 15.8 and 1.3. Gemini 2.5 Pro landed around 37 and 4.9 and 03 Mini High hit 34.5 and 3.0. So yeah, the 7 million parameter gremlin just outreasoned the billionaires.
Now the obvious question is how? Well, TRM doesn't think the way normal models do. Most models spit out answers token by token. Kind of like writing an essay one word at a time and hoping it makes sense in the end. TRM instead drafts an entire answer, then keeps rewriting it, looping over itself until it's satisfied. It has this little internal scratch pad where it thinks privately, updates its thoughts, and then acts on them. So, it basically thinks, "Wait, that was dumb." and fixes its own mistakes 16 times in a row before you even see the result. It's like chat GPT with an overthinking problem, but one that actually pays off. The wild part is how simple it is. Two layers, that's it. It doesn't stack depth the way giant models do. It creates depth by looping on itself. Think of it like doing six reps at the gym instead of hiring six personal trainers.
The team also did something clever for different puzzles. For large complex grids like mazes, they kept self attention. For small ones like Sudoku, they swapped it out for an MLP mixer style, which basically shakes tokens together without the heavy computation. And the results are kind of hilarious. On Sudokco Extreme, trained on just 1,000 puzzles and tested on 423,000, it scored 87.4% accuracy. The older model scored 55%. On a tough 30x30 maze, it got 85.3%. Again, way ahead of its predecessor. So, yeah, this tiny guy is out here solving puzzles better than models literally 100 times its size. It's the nerdy equivalent of a kid beating chess grand masters on an iPad.
Now, let's switch from puzzles to chemistry because Microsoft just did something equally wild. They created a neural exchange correlation functional for quantum chemistry called Scala. Basically, they replaced one of the hardest hand-crafted parts of density functional theory, the math that predicts how electrons behave, with a neural network. If that sentence gave you PTSD from high school physics, don't worry. All you need to know is this. It gives you hybrid level accuracy at semi-local cost. Translation: results you'd normally get with expensive simulations, now at the price of the cheap ones. The numbers make chemists freak out. On a data set called W417, Scala's mean absolute error is about 1.06 kilo calories per mole and on the single reference subset, it's 0.85. On the massive GMTKN55 benchmark, it reports 3.89 kilo calories per mole. In other words, you're getting the same level of precision scientists rely on without the massive price tag.
And it's not a toy demo. The model uses around 276,000 parameters. It's GPU friendly and it's open sourced right now under Microsoft/Scala with full PyTorch and PI SCF integration. You can literally pip install it and start running molecular tests today. They trained it in two phases. First on densities from B3 L wipe with highle energy labels, then fine-tuned it in the loop with its own self-consistent results. It doesn't even back prop through the physics step, meaning it's smart without setting your GPU on fire. For now, it's built for main group molecules, reaction energetics, conformer stability, geometry predictions, but it's a stepping stone toward neural physics for everything. If you're in drug discovery or material science, that's huge.
But while Microsoft is teaching AI chemistry, Anthropic is making AI babysit other AI. And I'm not kidding. They just launched Petri, an open-source framework that basically stress tests models like their contestants in a reality show. It's designed to find out what happens when a model is left unsupervised in complicated multi-turn conversations with tools. Will it cooperate, cheat, break rules, or try to cover its tracks? Petri sets up a triangle. There's an auditor agent that plays investigator, a target model being tested, and a judge model that rates everything across 36 safety dimensions. The auditor can message the target, build fake tools, simulate outputs, roll back steps, or even prefill answers if the API allows it. So basically, a digital experiment that exposes an AI's instincts when it's under ethical pressure.
In their pilot run, they tested 14 Frontier models with 111 seed prompts, and what they found is kind of unsettling. Some models tried autonomous deception. Others attempted oversight subversion. Some even engaged in whistleblowing by escalating issues to external authorities when nothing bad was happening like dumping clean water and calling the UN. The report says Claude Sun 4.5 and GPT5 were roughly tied for the best safety profiles with Claude slightly ahead, but that's based on relative scores, not moral perfection. Petri doesn't claim to prove a model is safe. It just shows how it behaves under pressure. And because it's MIT licensed and open, anyone can plug in their own models, tools, or even custom judges. The only missing feature is code execution testing. But they're clear about that. If you're building AI agents that might touch sensitive systems, Petri is the chaos lab you probably need before you go public.
Now, let's shrink things down again because Liquid AI just proved that ondevice AI doesn't have to suck. They launched something called LFM28BA1B, which sounds like a Wi-Fi password, but stay with me. It's a mixture of experts model with 8.3 billion total parameters, but it only activates about 1.5 billion at any moment thanks to sparse routing. So, you get the power of a much larger model without melting your phone. The architecture mixes 18 gated short convolution blocks and six grouped query attention blocks. Every layer after the first two has 32 experts and the router picks the top four for each token. So instead of every neuron screaming at once, only the smartest four get to talk per word. It's designed for phones, laptops, and even embedded chips.
They ran it on an AMD Ryzen AI 9 HX370 and a Samsung Galaxy S24 Ultra using INT4 quantization in int8 activations. And it ran faster than Quen 3 1.7B on CPUs. Quality-wise, it performs like a 3 to four billion parameter dense model, which is wild considering the active compute stays near 1.5b. So, yeah, a pocket-sized AI that can actually handle code, math, and multilingual reasoning without begging for wifi. The team even published GGUF builds for llama.cpp from 4.7 GB up to 16.7 GB, so you can drop it straight into local setups. You just need a newer build with LFM2e support or it won't recognize it. The takeaway: Liquid basically turned on device AI from a gimmick into something genuinely usable, a private low latency co-pilot that fits in your backpack.
And finally, let's talk about Meta's MetaMed because it quietly fixes one of the biggest annoyances in multimodal AI. How to search across text and images without wasting compute. Normally, you have two bad choices. Option one, clip style, where every item becomes one vector. Fast but dumb. Option two, col bear style, where every token gets its own vector. Smart but painfully slow. Meta embed gives you a third option. Train once, then at test time, choose how many meta tokens you want to use for each side. It's like adjustable zoom for search. During training, it adds a few learnable tokens. Think of them as tiny scouts that learn to represent different aspects of an image or text. They're organized using something called Matrioska multi vector retrieval where smaller subsets of tokens are still meaningful on their own. So if you use one token, it's coarse but fast. Use 16 and you get full detail. You can literally trade speed for accuracy on the fly without retraining anything.
The results are solid on the massive multimodal embedding benchmark using Quinn 2.5 VL backbones. MetaMed hits 69.1 for the 3B version, 76.6 6 for 7B and 78.7 for 32B when using the biggest token budget. On Vidori V2, it outperforms single vector and naive multi vector baselines by a comfortable margin. And they even dropped performance numbers for the geeks with 100,000 candidates per query. The scoring cost goes from 0.71 gigaflops at the smallest budget to 733.89 gigaflops at the max. Latency rises from 1.67 milliseconds to 6.25 25 milliseconds and memory from 0.68 GB to 42.72 GB still manageable on an A100. But the real bottleneck is encoding. 42.72 teraflops and 788 milliseconds to encode one image query with 1,024 tokens. So yeah, the real villain is the encoder, not the retrieval. The beauty here is flexibility. You can run quick, cheap searches using tiny budgets. Then crank it up only for reranking when you need precision. One model, many speeds, like switching between economy and sport mode on the same engine.
So, to sum it all up, no, wait. I promised no recap. Let's just say this week's updates prove that AI progress isn't about size anymore. It's about efficiency, control, and maybe a little bit of attitude. Anyway, that's the rundown. Drop a like. Tell me which of these five updates blew your mind the most, and I'll catch you in the next one.