Share this @internewscast.com
AI startup Thinking Machine Labs releases its first-ever product called Tinker and here’s what it’s all about.
getty
In today’s column, I examine a newly announced AI product and service named Tinker that is being brought out by Thinking Machines Labs. TML is a startup company that was co-founded by the famed Mira Murati (former CTO at OpenAI). The nascent firm has been widely lauded as a shining star of top-grade AI startups.
Note that Tinker is the first-ever product from TML. Expectations were high. Everyone was excitedly sitting on the edge of their seats. What would this amazing startup put together? Could their first product knock our socks off? Might they reveal some incredible, earth-shattering AI capacity, perhaps something showcasing that AGI is just around the corner?
Prepare yourself for what might be a bit of a letdown.
Though Tinker is undoubtedly useful, providing a handy means to tune LLMs, there is a bit of head-scratching taking place. This is a multi-billion-dollar startup. They have some of the greatest AI developers in the world on their team. Instead of an eye-popping first product, we are greeted with a rather pedestrian capability that is certainly a nice-to-have, but not even close to hitting the ball out of the ballpark. Some are dismayed. Some think we are being trolled. Others insist that this is a fine way to get things underway and that we will indubitably be overjoyed once the second or third product gets released someday.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
The AI Startup Of All AI Startups
In case you aren’t already familiar with Thinking Machines Lab, I will go ahead and quickly get you up to speed. If you already know about the background of the notable startup, you are welcome to skip down to the rundown of what Tinker is all about.
The story about TML goes like this. In February 2025, Mira Murati co-founded the company, having left OpenAI, where she had been the CTO and, for a short while, served as the interim CEO during the Sam Altman commotion in November 2023 (see my coverage at the link here). By Spring 2025, TML had established a team of about thirty top-notch AI developers and researchers. Big names. Impressive talent.
The money came next. In June 2025, a seed round raised a whopping $2 billion at a reported $10 billion valuation. The investment process was led by Andreessen Horowitz (A16Z), and prominent investors included Nvidia, AMD, Cisco, and others. A ton of bucks. Lots of cash to provide a suitable runway to pursue something of immense magnitude and a presumed game-changer toward the advancement of AI.
The core mantra for TML is ardently stated on their website, namely, aiming to build a world-class team to solve the most challenging problems in artificial intelligence. Also, the overarching high-minded purpose or mission is: “Thinking Machines Lab is a research and product company making artificial intelligence more understandable, customizable, and capable for everyone.”
It is a grand convergence of talent, money, and a passionate desire to move the needle forward on the challenges of advancing AI. Goosebumps are the order of the day. I assume you can readily discern why there are heightened expectations and a watchful eye on what TML might accomplish.
Tinker Is Announced
On October 1, 2025, TML revealed its first-ever product. I say product, but it really is primarily a service that connects you to their devised software product. The purpose of Tinker is to help AI developers, AI researchers, and other AI wonks with fine-tuning a generative AI model.
In brief, if someone wants to boost a large language model and fine-tune it to a specialized domain or otherwise tailor-fit the LLM, they could use Tinker to do so. Anyone doing this would pretty much have to be relatively skilled in AI. The interface to access Tinker is via an API (application programming interface). You provide Tinker with the LLM that you want to fine-tune, you use the API to activate your desired fine-tuning, and you can then download the result and proceed with your newly augmented AI.
According to a TML posting entitled “Tinker: A Training API For Researchers and Developers,” here are the key elements regarding Tinker (excerpts):
- “Today, we are launching Tinker, a flexible API for fine-tuning language models.”
- “Tinker is a managed service that runs on our internal clusters and training infrastructure.”
- “We handle scheduling, resource allocation, and failure recovery. This allows you to get small or large runs started immediately, without worrying about managing infrastructure.”
- “We use LoRA so that we can share the same pool of compute between multiple training runs, lowering costs.”
- “Tinker is now in private beta for researchers and developers. Tinker will be free to start.”
The notion of and the concerted effort underlying Tinker have been a matter of great stealth. It has been kept a tightly held secret by TML. Mum is the word. Now, after eagerly awaiting this moment, the secret has finally been revealed. Now we know. The world at large knows.
For the moment, Tinker is only going to be selectively used by those who sign up on a waitlist and get chosen to participate in the private beta. It is currently available for just a few selected LLM open models. Those managing to get the privilege to use Tinker will be able to do so initially for free. Whether the free avenue will always be the case remains to be seen. Naturally, the assumption is that at some point a fee arrangement will be identified, and the product/service will be more widely made available.
The Problem And The Solution
You might be wondering what the problem is that Tinker aims to solve. I mentioned above that it has something to do with fine-tuning LLMs. There is more to this. I will gingerly walk you through some of the details to introduce you to a nagging problem that is well-known by AI insiders.
First, we need to get under the hood about foundational components of modern-era LLMs.
LLMs typically have at their roots a large data structure that is referred to as an artificial neural network or ANN. This is different from a true neural network, such as the biological one that resides inside your head (smarmily coined as wetware by AI devotees). An ANN is a mathematical and computational data structure that contains lots and lots of numbers. Those numbers are determined when the LLM is initially data trained. You feed massive volumes of data from the Internet into algorithms that pattern-match on the data and then calculably represent the patterns via numbers within the ANN.
For more about how ANNs and LLMs are constructed and function, see my coverage at the link here.
At a 30,000-foot level, let’s loosely refer to those numbers as being parameters. How many parameters might a viable LLM have? Well, it depends. Meta’s Llama is an open LLM that can be set up in various sizes, ranging from 1 billion to perhaps 2 trillion parameters. A closed LLM is a proprietary AI, and we don’t necessarily know how many parameters exist since the AI maker might not want to publicly say what it is. For example, OpenAI’s proprietary or closed LLM of GPT-5 might have around 600 billion parameters, and the earlier predecessor GPT-3 perhaps has 175 billion parameters. Those are guesses.
Scaling Forever Or Hit The Wall
The massive size of the number of parameters is the wizardry that has made generative AI so fluent and compelling to use. Whereas early versions of LLMs are tiny in comparison and did not do much, it was discovered that scale matters. Scale matters a whole heck of a lot.
As we keep scaling up the size and increasing the number of parameters, so far, the results continue to be impressive. Whether this will continue indefinitely is one of the biggest debates in the AI field. One argument is that we just need to keep throwing more computer memory and processing speed (plus time) at LLMs. This is going to get us to artificial general intelligence and possibly artificial superintelligence, the hoped-for AGI and ASI.
Others express doubt about this assumption. A worry is that at some juncture, we will hit a wall. No matter how many more parameters you include, the LLM isn’t going to get much better. A diminishing point of returns is going to arise. Based on that concern, some believe we need to pursue other alternative architectures and completely reimagine how we devise AI. See my analysis of this heady topic at the link here.
The Problem Entails Fine-Tuning
You are ready now for the big reveal.
Suppose that we consume gobs and gobs of computer processing during initial data training and end up with our massive-sized LLM. It was pattern-matched on a wide swath of the Internet. Nice. We have a generalized AI that can do the usual stuff of answering everyday questions.
But we want our LLM to also cover a specific domain, perhaps being fine-tuned to perform highly specialized medical diagnoses. The initially performed widespread scan didn’t happen to encompass the intricacies of this specific medical subfield. Darn.
What can we do to get the LLM immersed in the subfield?
A brute force approach would be to utterly retrain the entire model. This might require revisiting and recalculating all the billions of parameters. The amount of processing time and consumption of GPUs could be enormous. Is it worth it? The issue arises as to the cost associated with the fine-tuning versus what benefits we hope to gain by including the desired subfield.
You might be wondering whether we could try to do something other than a brute force attack. Yes, you have hit the nail on the head. There are other ways to fine-tune without having to entirely revisit the whole model (see some examples that I identify in my coverage at the link here).
Shortcuts have indeed been discovered.
The Magnificent LoRA
A shortcut method that many in the AI field like to use is known as “low-rank adaptation” and is colloquially referred to as LoRA.
LoRA is considered a viable shortcut technique for fine-tuning an LLM. Usually, the fine-tuning occurs quickly and inexpensively compared to the brute force approach. This entails a clever trick. Rather than visiting all or most of the parameters, we can attach some new pieces to the ANN and only have to visit a much smaller number of parameters. And, to a certain extent, it accomplishes the inclusion of the subfield, though it isn’t a cure-all, and important tradeoffs are involved.
LoRA is a highly popular go-to by AI devotees.
The big reveal then is this: Tinker is an implementation of LoRA (you might have keenly observed that the above TML notable elements mentioned LoRA, kudos for catching that). An AI developer or AI researcher can use an off-the-shelf LoRA that is provided by TML and do so without the usual hassles involved. You don’t need to code your own LoRA. You don’t need to find an existing LoRA and figure out whether it works properly. Etc.
TML allows you to access an API to get to the LoRA, which, presumably, they have extensively devised and tested (you can seemingly rely on it), plus, conveniently, it runs on their hardware and architecture. It is both a software product and a service offering.
LoRA Is Old News In A New Bottle
You might be tempted to think that LoRA is something newly discovered in the last several months, and maybe we are only now just opting to use the technique. Nope, that would be an incorrect assumption. The history of LoRA goes back to at least 2022, and some years earlier, regarding when it was first thought of, formulated, and mindfully worked out.
In a now-classic research article from 2022, entitled “LoRA: Low-Rank Adaptation Of Large Language Models” by Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, Conference Proceeding of ICLR 2022, the LoRA technique was formally anointed:
- “Many applications in natural language processing rely on adapting one large-scale, pre-trained language model to multiple downstream applications. Such adaptation is usually done via fine-tuning, which updates all the parameters of the pre-trained model.”
- “Fine-tuning enormous language models is prohibitively expensive in terms of the hardware required and the storage/switching cost for hosting independent instances for different tasks.”
- “We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable parameters for downstream tasks.”
- “Compared to GPT-3 175B fine-tuned with Adam, LoRA can reduce the number of trainable parameters by a factor of 10,000 and the GPU memory requirement by a factor of 3.”
The empirical work in that classic paper provided impetus for further extending the LoRA technique. All sorts of variations of LoRA exist today. You can get some of them for free, and others you must pay to use. The marketplace of LoRA software is a bit of a narrow niche. It is perhaps evident that the need for LoRA is sought by a somewhat narrow set of AI devotees.
LoRA isn’t an especially booming, widespread, or in-demand capability by the public at large. It is a proverbial inside-the-beltway AI-nerdish consideration. AI insiders look around, find a LoRA that meets their needs, and use it when they are asked to do fine-tuning in certain circumstances.
Analysis Of Tinker Macroscopically
You are primed for diving into the AI-insider head-scratching.
I noted at the opening that many had anticipated that TML would come out of the gate with something awesome and dazzling. It would be new. It would be so innovative that our jaws would drop. It would be a clear indicator that AGI can be had, and possibly ASI too.
What we got instead is an instance of LoRA that assuredly must be top-notch, and that you can conveniently use as a service. Sure, that’s great. Does it imply or signal that we are flying at lightning speed toward AGI? Sorry, no. Is it a breakthrough in the advancement of AI? Not really, even if their LoRA is souped up, you would be hard-pressed to contend it was an entirely eye-popping out-of-thin-air innovation.
We know LoRA. It is an old friend. Old friends are welcome, but it would be exciting to meet someone new.
Explanations Speculations Galore
AI enthusiasts are trying to figure out what this all means.
Were they perhaps working on a LoRA offering when they were first founded as a company, and this happens to be the first fully cooked meal ready to be served? They may have decided to get this out the door and showcase what they can accomplish. Good for them.
Another possibility is this. Some suggest that they have assembled a talented group of heads-down AI gurus and thus are preoccupied with interesting but somewhat arcane technicalities. Perhaps marketers weren’t especially incorporated into the mix. The aim of looking at the marketplace and trying to find a huge money-making mover-and-shaker of a new product wasn’t at the heart of their work at hand. The techies knew the fine-tuning problem existed; they had in their minds a means of super-charging LoRA, and they have victoriously met that goal. Period, end of story.
Whoa, some exhort, calm down, and don’t undervalue what Tinker provides.
You are gauging this high-stepping company by a perspective that is unfair and premature. They are showcasing their prowess. Furthermore, think about what this LoRA might unleash. It could be that everyone and their brother or sister will now seek to fine-tune LLMs. You have to see the forest for the trees. Whereas LoRA had been a challenge to adopt, it can now be used at the drop of a hat.
LLMs will be rapidly fine-tuned and encompass zillions of highly desired domains and subdomains. We are unlocking the potential of generalized AI. The next thing we know, if you need an LLM that does this or that, voila, it either has already been fine-tuned via the use of Tinker, or it can be quickly and inexpensively fine-tuned by invoking Tinker.
Boom, drop the mic.
We Live In Interesting Times
Ponder a reflective analogy of the situation.
Imagine that a new filmmaking company was auspiciously formed. It brought together the best movie directors, the best actors and actresses, and was otherwise jam-packed with the best talent in the entertainment business. The startup got funded to the tune of billions of dollars.
What kind of movie would we expect this awesome company to release first?
A blockbuster, of course.
Suppose that instead, they released a praiseworthy artsy film that was well-devised, cleverly put together, but appealed primarily to a relatively small segment of the marketplace. Eyebrows would undoubtedly be raised. Admiration for the fine artisan touches would be gushingly expressed. Meanwhile, the anticipation for a blockbuster would still permeate the ether surrounding this grand filmmaker enterprise.
As the famous English poet Lady Mary Montgomerie once said: “Great things come to those who wait.” I am eager to see what else TML conjures up and will endeavor to give whatever it is a discerning close-in look.
You’ll be the first to know.