Claude Code is doing pretty well in my experience :) I've built a tool in our CI environment that reads Jira tickets, files GitHub PRs, etc. automatically. Great for one-shotting bugs, and it's only getting better.
Integrations are nice, but the superpower is having an AI smart enough to operate a computer/keyboard/mouse so it can do anything without the cooperation/consent of the service being used.
Lots of people are making moves in this space (including Anthropic), but nothing has broken through to the mainstream.
Why can't one set up a prompt, test it against a file, then once it is working, apply it to each file in a folder in a batch process which then provides the output as a single collective file?
I've just done something similar with Claude Desktop and its built-in MCP servers.
The limits are still buggy responses - Claude often gets stuck in a useless loop if you overfeed it with files - and lack of consistency. Sometimes hand-holding is needed to get the result you want. And it's slow.
But when it works it's amazing. If the issues and limitations were solved, this would be a complete game changer.
We're starting to get somewhat self-generating automation and complex agenting, with access to all of the world's public APIs and search resources, controlled by natural language.
I can't see the edges of what could be possible with this. It's limited and clunky for now, but the potential is astonishing - at least as radical an invention as the web was.
I've been using Claude Desktop with built-in File MCP to run operations on local files. Sometimes it will do things directly but usually it will write a Python script. For example: combine multiple.md files into one or organize photos into folders.
I also use this method for doing code prototyping by giving it the path to files in the local working copy of my repo. Really cool to see it make changes in a vite project and it just hot reloads. Then I make tweaks or commit changes as usual.
I think all the retail LLM's are working to broaden the available context, but in most practical use-cases it's having the ability to minimize and filter the context that would produce the most value. Even a single PDF with too many similar datapoints leads to confusion in output. They need to switch gears from the high growth, "every thing is possible and available" narrative, to one that narrows the scope. The "hallucination" gap is widening with more context, not shrinking.
That's a tough pill to swallow when your company valuation is a $62B based on the premise that you're building a bot capable of transcendent thought, ready to disrupt every vertical in existence.
Tackling individual use-cases is supposed to be something for third party "ecosystem" companies to go after, not the mothership itself.
This has been my experience as well. The moment you turn internet access on, Kagi Assistant starts outputting garbage. Turn it off and you're all good.
Definitely my experience. I manage context like a hawk, be it with Claude-as-Google-replacement or LLM integrations into systems. Too little and the results are off. Too much and the results are off.
Not sure what Anthropic and co can do about that, but integrations feel like a step in the wrong direction. Whenever I've tried tool use, it was orders of magnitude more expensive and generally inferior to a simple model call with curated context from SerpApi and such.
Couldn't agree more. I wish all major model makers would build tools into their proprietary UIs to "summarize contents and start a new conversation with that base". My biggest slowdown with working with LLMs while coding is moving my conversation to a new thread because context limit is hit (Claude) or the coherent-thought threshold is exceeded (Gemini).
I never use any web interfaces, just hooked up gptel (an Emacs package) to Claude's API and a few others I regularly use, and I just have a buffer with the entire conversation. I can modify it as needed, spawn a fresh one quickly etc. There's also features to add files and individual snippets, but I usually manage it all in a single buffer. It's a powerful text editor, so efficient text editing is a given.
I bet there are better / less arcane tools, but I think powerful and fast mechanisms for managing context are key and for me, that's really just powerful text editing features.
you hit the nail on the head. my experience with prompting LLMs is that providing extra context that isn’t explicitly needed leads to “distracted” outputs
I mean, to be honest, they gotta do both to achieve what they’re aiming for.
A truly useful AI assistant has context on my last 100,000 emails - and also recalls the details of each individual one perfectly, without confusion or hallucination.
Obviously I’m setting a high bar here; I guess what I’m saying is “yes, and”
This is my concern as well. How successful is it in selecting the correct tool out of hundreds or thousands?
Different to what this integration is pushing, the LLMs usage in production based products where high accuracy is a requirement (99%), you have to give a very limited tool set to get any degree of success.
Had been planning a custom mcp for our orgs’ jira.
I’m a bit skeptical that it’s gonna work out of the box because of the amount of custom fields that seem to be involved to make successful API requests in our case.
But I would welcome, not having to solve this problem. Jira’s interface is among the worst of all the ticket tracking applications I have encountered.
But, I have found using a LM conversation paired within enough context about what is involved for successful POSTs against the API allow me to create update and relate issues via curl.
It’s begging for a chat based LLM solution like this. I’d just prefer the underlying model not be locked to a vendor.
Atlassian should be solving this for its customers.
I'm familiar with that MCP and was planning to build on top of it.
I hadn't realized but the new integration seems to actually just be an official, closed-source MCP produced *by* Atlassian.
sooperset's MCP is MIT licensed, so I wonder how much of the Atlassian edition is just a lift of that.
There's a comment [1] on the actual integration page asking about custom fields, which I think is possibly a big issue.
At first I thought the open-source version would get crushed by an actual Atlassian release, but not if Atlassian doesn't offer all the support for it to work really well no matter what customizations are fitted into each instance.
My hypothesis is that it takes custom code to make this work, and using the off-the-shelf for Jira won't work. Hoping to be proven wrong though, as it would be less work for me on that front.
I feel dumb but how do you actually add Zapier or Confluence or custom MCP on the web version of claude? I only see it for Drive/Gmail/Github. Is it zoned/slow release?
The leap frogging at this point is getting insane (in a good way, I guess?). The amount of time each state of the art feature gets before it's supplanted is a few weeks at this point.
LLMs were always a fun novelty for me until OpenAI DeepResearch which started to actually come up with useful results on more complex programming questions (where I needed to write all the code by hand but had to pull together lots of different libraries and APIs), but it was limited to 10/month for the cheaper plan. Then Google Deep Research upgraded to 2.5 Pro and with paid usage limits of 20/day, which allowed me to just throw everything at it to the point where I'm still working through reports that are a week or more old. Oh and it searched up to 400 sources at a time, significantly more than OpenAI which made it quite useful in historical research like identifying first edition copies of books.
Now Claude is releasing the same research feature with integrations (excited to check out the Cloudflare MCP auth solution and hoping Val.town gets something similar), and a run time of up to 45 minutes. The pace of change was overwhelming half a year ago, now it's just getting ridiculous.
I agree with your overall message - rapid growth appears to encourage competition and forces companies to put their best foot forward.
However, unfortunately, I cannot shower much praise on Claude 3.7. And if you (or anyone) asks why - 3.7 seems much better than 3.5, surely? - Then I’m moderately sure that you use Claude much more for coding than for any kind of conversation. In my opinion, even 3.5 Haiku (which is available for free during high loads) is better than 3.7 Sonnet.
Here’s a simple test. Try asking 3.7 to intuitively explain anything technical - say, mass dominated vs spring dominated oscillations. I’m a mechanical engineer who studied this stuff and I could not understand 3.7’s analogies.
I understand that coders are the largest single group of Claude’s users, but Claude went from being my most used app to being used only after both chatgpt and Gemini, something that I absolutely regret.
I use Claude mostly for coding/technical things and something about 3.7 does not feel like an upgrade. I haven't gone back to 3.5 (mostly started using Gemini Pro 2.5 instead).
I haven't been able to use Claude research yet (it's not rolled out to the Pro tier) but o1 -> o3 deep research was a massive jump IMHO. It still isn't perfect but o1 would often give me trash results but o3 deep research actually starts to be useful.
3.5->3.7 (even with extended thinking) felt like a nothingburger.
3.7 did score higher in coding benchmarks but in practice 3.5 is much better at coding. 3.7 ignores instructions and does things you didn't ask it to do.
I think it just does that to eat up your token quota and get you to upgrade.
Like, ask it a simple question and it comes up with a full repo, complete with a README and a Makefile, when all you wanted to know was how efficient a particular algorithm would be in the included code.
Can't wait until the add research to the Pro plan because, you know, I have questions...
> I think it just does that to eat up your token quota and get you to upgrade.
If you pay for a subscription then they don’t have an incentive to use more tokens for the same answer.
It’s definitely because feedback from people has “taught” it that more boilerplate is better. It’s the same reason ChatGPT is annoyingly complementary.
Gemini 2.5 Pro has solved problems that Claude 3.7 cannot, so I use it for the hard stuff.
But Gemini is at least as overactive as Claude, sometimes even more overactive when it comes to something like comment spam.
Of course, this can be fixed with prompting. And sometimes it feels sheepish complaining about the machine god doing most of my chore work that didn't even exist a couple years ago.
> My current hypothesis: the more familiar you are with a topic the worse the results from any LLM.
That's not really true, since your prompts are also getting better. Better input leads to better output remains true, even with LLMs (when you see it as a tool).
Being more familiar with the topic definitely doesn't always make your prompts better. For a lot of things it doesn't really change (explain X, compare X and Y...) - and this is what is being discussed it. For giving "building" instructions (like writing code) it helps a bit, but even if you know exactly what you want it to write, getting it to do that is pretty much trial and errror (too much detail makes it follow word-for-word and produce bad code, too little and it misses important parts or makes dumb mistakes).
The opposite may be true, the more effective the model the lazier the prompting as it can seemingly handle not being micromanaged as with earlier versions.
That is certainly the case in niche topics where published information is lacking, or needs common sense to synthesize proper outputs [1].
However in this specific example, I don't remember if it was chatgpt or gemini or 3.5 Haiku but the other(s) explained it well enough. I think I re-asked 3.5 Haiku at a later point of time, and to my complete non-surprise, it gave an answer that was quite decent.
1 - For example, the field of DIY audio - which was funnily enough the source of my question. I'm no speaker designer, but combining creativity with engineering basics/rules of thumb seems to be something LLms struggle with terribly. Ask them to design a speaker and they come up with the most vanilla, tired, textbook design - despite several existing market products that are already so much ahead/innovative.
I'm confident that if you asked an LLM an identical question for which there is more discourse - eg make an interesting/innovative phone - you'd get relatively much better results.
Yes, it’s particularly bad when the information found on the web is flawed.
For example, I’m not a domain expert, but I was looking for an RC motor for a toy project and OpenAI had happily tried to source a few, with Deep Research. Only the best candidate it had picked contained an obvious typo in the motor spec (68 grams instead of 680 grams), which is just impossible for a motor of specified dimensions.
Right, but LLMs are also consuming AWS product documentation and Terraform language docs, some things I have read a lot of and they’re often badly wrong on things from both of those domains, which are really easy for me to spot.
This isn’t just “shit in, shit out”. Hallucination is real and still problematic.
I had it generate a baseball lineup the other day, it printed out a list of the
13 kids names, then said (12 players). Just straight up miscounted what it was doing, throwing a wrench to everything else it was doing beyond that point.
It is like this with expert humans too. Which is why, no matter what, we will continue to require expert humans not just "in the loop" but as the critical cogs that are the loop itself, just as it as always been. However, this time around those people will have AI augmentation, and be intellectually athletes of a nature our civilization has never seen.
The more familiar you are with the state of “Jira hygiene” in the megacorp environment, the less hope you have that LLMs will be able to make sense of things.
That said, the “AI all the things” mandates could be the lever that ultimately accomplishes what 100+ PjMs couldn’t - making people write issues as if they really mattered. Because garbage in, garbage out.
Seems clear to me that Claude 3.7 suffers from overfitting, probably due to Anthropic seeing that 3.5 was a smash hit in the LLM coding space and deciding their North star for 3.7 should be coding benchmarks (which, like all benchmarks, do not properly capture the process of real-world coding).
If it was actually good they would've named it 4.0, the fact that they went from 3.5 to 3.7 (weird jump) speaks volumes imo.
The numbering jump is because there was "Claude 3.5" and then "Claude 3.5 (new)" and they decided to retroactively stop the madness and rename the later to 3.6 (which is what everyone was calling it anyway).
Plateauing overall but apparently you can gain in certain directions while you lose on some. I've written an article a while back that current models are not that far from GPT-3.5: https://omarabid.com/gpt3-now
3.7 is definitively better at coding but you feel it lost a bit of maneuverability at other domains. For someone who wants code generated, it doesn't matter but I've found myself using DeepSeek first and then getting code output by 3.7.
None of those reports are any good though. Maybe for shallow research, but I haven't found them deep. Can you share what kind of research you have been trying there where it has done a great job of actual deep research.
Deep Research hasn't really been that good for me. Maybe I'm just using it wrong?
Example: I want the precipitation in mm and monthly high and low temperature in C for the top 250 most populous cities in North America.
To me, this prompt seems like a pretty anodyne and obvious task for Deep Research. It's long, tedious, but mostly coming from well structured data sources (wikipedia) across two languages at most.
But when I put this in to any of the various models, I mostly get back ways to go and find that data myself. Like, I know how to look at Wikipedia, it's that I don't want to comb through 250 pages manually or try to write a script to handle all the HTML boxes. I want the LLM/model to do this days long tedious task for me.
The funny thing is that if your request only needed the top 100's temperature or the top 33's precipitation, it could just read "List of cities by average temperature" or "List of cities by average precipitation" and that would be it, but the top 250 requires reading 184x more pages.
My perspective on this is that if Deep Research can't do something, you should do it yourself and put the results on the internet. It'll help other humans and AIs trying to do the same task.
The project requires the full list of every known city in the western hemisphere and also Japan, Korea, and Taiwan. But that dataset is just maddeningly large, if it is possible at all. Like, I expect it to take me years, as I have to do a lot of translations. So, I figured that I'd be nice and just as for the top 250 for the various models.
There's a lot more data that we're trying to get too and I'm hoping that I can get approval to post it as its a work thing.
If you have the data, but need to parse all of it, couldn’t you upload it to your LLM of choice (with a large enough context window) and have it finish your project?
Well remember listing/ranking things are structurally hard for these models because you have to keep track of what it has listed and what it hasn't, etc.
Sounds like the you're having it conduct research and then solve the Knapsack problem for you on the collected data. We should do the same for the traveling salesman one.
How do you validate its results in that scenario? Just take its word for it?
Ahh, no. We'll be doing more research on the data once we have it. Things like ranking and averages and distributions on the data will come later, but first we just need it to begin with.
That's actually not what deep research is for, although you can obviously use it however you like. Your query is just raw data collection—not research. Deep research is about exploring a topic primarily with academic and other high-quality sources. It's a starting point for your own research. Deep research creates a summary report in ~10 min from more sources than you could probably read in a month, and then you can steer the conversation from there. Alternatively, you can just use deep research's sources as a reading list for yourself so you can do your own analysis.
I think we have very different definitions of the word 'research' then.
I'd say that what you're saying is 'synthesis'. The 'Intro/Discussion' sections of a journal article.
For me, 'research' means the work of going through and getting all the data in the first place. Like, going out and collecting dino bones in the hot sun, measuring all the soil samples, etc. - that is research. For me, asking these models to go collate some webpages, I mean, you spend the first weeks of a summer undergrad's time to go do this kid of thing to get them used to the file systems and spruce up their organization skills, see where they are at. Writing the paper up, that's part of research sure, but not the hard part that really matters.
Agreed—we're working with different definitions of "research". The deep research products from OpenAI, Google Gemini, and Perplexity seem to be more aligned with my definition of research if that helps you gain more utility from them.
It's excellent at producing short literature reviews on open access papers and data. It has no sense of judgment, trusting most sources unless instructed otherwise.
Gemini's Deep Research is very good at discriminating between sources though, in my experience (haven't tried Claude or Perplexity). It finds really obscure but very relevant documents that don't even show up in Google Search for the same queries. It also discounts results that are otherwise irrelevant or very low-value from the final report. But again, it is just a starting point as the generated report is too short, and I make sure to check all the references it gives once again. But that's where I find its value.
The failure mode is that people unfamiliar with a subject aren't able to distinguish careful analysis from bullshit. However the second failure mode where someone pointing that out is assumed to be calling people stupid is a longstanding wetware bug.
Out of curiosity - can you give any examples of the programming questions you are using deep research on? I’m having a hard time thinking of how it would be helpful and could use the inspiration.
Easy, any research task that will take you 5 minutes to complete it's worth firing off a Deep Research request while you work on something else in parallel.
I use it a lot when documentation is vague or outdated. When Gemini/o3 can't figure something out after 2 tries. When I am working with a service/API/framework/whatever that I am very unfamiliar with and I don't even know what to Google search.
I recently asked Chrome to show me how to apply the Knuth-Bendix completion procedure to propositional logic, and I had already formed my own thoughts about how to proceed (I'm building a rewrite system that does automated reasoning).
The response convinced me that I'm not a total idiot.
I'm not an academic and I'm often wrong about theory so the validation is really useful to me.
That’s a perfect example of LLMs providing epistemic scaffolding — not just giving you answers, but helping you check your footing as you explore unfamiliar territory. Especially valuable when you’re reasoning through something structurally complex like rewrite systems or proof strategies. Sometimes just seeing your internal model reflected back (or gently corrected) is enough to keep you moving.
I've been using it for pre scoping things I have no idea about and rapidly iterating by refeeding it a version with guard rails and conditions from previous chats.
Like I wanted to scope how to build a home made TrueNAS Scale unit, it helped me with a avoiding pitfalls like knowing that I needed two GPUs minimum to run the OS and local llms, and speed up config for a CLI back up of my Dropbox locally(it told me to use the right filesystem format over ZFS to make Dropbox client work).
It has researched how I can structure my web app for building payment system on the web(something I knew nothing about) to writing small tools to talk to my document collection and index them into collections in Anki in one day.
You still need "human in the loop" because with simple tasks or some tasks that have lots of training material, models can one-shot answer and are like super good. But if the domain grows too complex, there are some not-so-obvious dependencies, or stuff that is in bleeding edge. Models fail pretty badly. So you need someone to split those complex tasks to more simpler familiar steps.
Calling some APIs is leap-frogging? You could do this with GPT-3, nothing has changed except it's branded under a new name and tries to establish a (flawed) standard.
If there was truly any innovation still happening in OpenAI, Anthropic, etc., they would be working on models only, not on side features that someone could already develop over a weekend.
In my previous team most of our oncall requests came from bug reports by customers on various tools that we owned, so to be able to work on random tools that my team owned was a nice change of pace / scenery compared to working on the same thing for 3 months uninterrupted.
Now I'm in a new team where 99% of our oncall tickets come from automated alarms and 80% of them are a subset of a few issues where the root-cause isn't easy to address but there is either nothing to actually do once investigated, or the fix is a one time process that is annoying to run, so the username isn't accurate anymore :)
I still like the change of pace though, 0 worries about sprint tasks or anything else for a week every few months.
i'm not an expert in either, but RAG is like dropping some 'useful' info into the prompt context, while fine tuning is more like a performing mix of retraining, appending re-interpretive model layers and/or brain surgery.
I'll leave it to you to guess which one is harder to do.
It’s significantly harder to get right, it’s a very big stepwise increase in technical complexity over in context learning/rag.
There are now some light versions of fine tuning that don’t update all the model weights but train a small adapter layer called Lora which is way more viable commercially atm in my opinion.
There were initial difficulties in finetuning that made it less appealing early on, and that's snowballed a bit into having more of a focus on RAG.
Some of the issues still exist, of course:
* Finetuning takes time and compute; for one-off queries using in-context learning is vastly more efficient (i.e., look it up with RAG).
* Early results with finetuning had trouble reliably memorizing information. We've got a much better idea of how to add information to a model now, though it takes more training data.
* Full finetuning is very VRAM intensive; optimizations like LoRA were initially good at transferring style and not content. Today, LoRA content training is viable but requires training code that supports it [1].
* If you need a very specific memorized result and it's costly to get it wrong, good RAG is pretty much always going to be more efficient, since it injects the exact text in context. (Bad RAG makes the problem worse, of course).
* Finetuning requires more technical knowledge: you've got to understand the hyperparameters, avoid underfitting and overfitting, evaluate the results, etc.
* Finetuning requires more data. RAG works with a handful datapoints; finetuning requires at least three orders of magnitude more data.
* Finetuning requires extra effort to avoid forgetting what the model already knows.
* RAG works pretty well when the task that you are trying to perform is well-represented in the training data.
* RAG works when you don't have direct control over the model (i.e., API use).
* You can't finetune most of the closed models.
* Big, general models have outperformed specialized models over the past couple of years; if it doesn't work now, just wait for OpenAI to make their next model better on your particular task.
On the other hand:
* Finetuning generalizes better.
* Finetuning has more influence on token distribution.
* Finetuning is better at learning new tasks that aren't as present in the pretraining data.
* Finetuning can change the style of output (e.g., instruction training).
* When finetuning pays off, it gives you a bigger moat (no one else has that particular model).
* You control which tasks you are optimizing for, without having to wait for other companies to maybe fix your problems for you.
* You can run a much smaller, faster specialized model because it's been optimized for your tasks.
* Finetuning + RAG outperforms just RAG. Not by a lot, admittedly, but there's some advantages.
Plus the RL Training for reasoning has been demonstrating unexpectedly effective improvements on relatively small amounts of data & compute.
So there's reasons to do both, but the larger investment that finetuning requires means that RAG has generally been more popular. In general, the past couple of years have been won by the bigger models scaling fast, but with finetuning difficulty dropping there is a bit more reason to do your own finetuning.
That said, for the moment the expertise + expense + time of finetuning makes it a tough business proposition if you don't have a very well-defined task to perform, a large dataset to leverage, or other way to get an advantage over the multi-billion dollar investment in the big models.
RAG is infinitely more accessible and cheaper than finetuning. But it is true that finetuning is getting severely overlooked in situations where it would outperform alternatives like RAG.
This assumes the team deploying the RAG-based solution has equal ability to either engineer a RAG-based system or to finetune an LLM. Those are different skillsets and even selecting which LLM should be finetuned is a complex question, let alone aligning it, deploying it, optimizing inference etc.
The budget question comes into play as well. Even if text is repetitively fed to the LLM, that might happen over a long enough time compared to finetuning which is a sort of capex that it is financially more accessible.
Now bear in mind, I'm a big proponent of finetuning where applicable and I try to raise awareness to the possibilities it opens. But one cannot deny RAG is a lot more accessible to teams which are likely developers / AI engineers compared to ML engineers/researchers.
You are certainly right, managed platforms make finetuning much easier. But managed/closed model finetuning is pretty limited and in fact should be named “distribution modeling” or something.
Results with this method are significantly more limited compared to all the power open-weight finetuning gives you (and the skillset needed in return).
And in either case don’t forget alignment and evals.
> Results with this method are significantly more limited compared to all the power open-weight finetuning gives you (and the skillset needed in return).
I am not sure I understand why you are so certain that finetuned top market models, built by top researchers will be significantly worse than whatever open source model you pick.
Nuance is hard. Binary choices are fast, comforting, and require less thought. Certainty feels safer than ambiguity — especially in conflict, where complexity threatens identity. And in most arenas (tech, media, politics), decisive hot takes get applause. Fence-sitters get ignored.
I worked at a startup where the CEO swore up and down that real-time fine-tuning was the future — that models would continuously update with company data. It sounded cool until you remember:
That’s not how LLMs work.
It’s not efficient.
It’s not flexible.
And it’s not even necessary — we already have RAG.
Pipedreams make good pitch decks. But they break when you hit production.
It's a fucking pipedream, this. That's not how LLMs work, it's not efficient, it's not useful (we have RAG for reference augmentation), and it’s not even desirable unless you want your model overfitting on stale, internal narratives every night.
Ongoing demo of integrations with Claude by a bunch of A-list companies: Linear, Stripe, Paypal, Intercom, etc.. It's live now on: https://www.youtube.com/watch?v=njBGqr-BU54
And what is the protocol for the interface between the GPU-based LLM and the API? How does the LLM signal to make a tool call? What mechanism does it use?
Because MCP isn’t an API it’s the protocol that defines how the LLM even calls the API in the first place. Without it, all you've got is a chat interface.
A lot of people misunderstand what is the role of MCP. It’s the signaling the LLM uses to reach out of its context window and do things.
Feed Claude the data willingly to learn more about human behavior they can’t scrape or obtain otherwise without consent? Hard pass. I’m not telling any AI any more about what it means to be a creative person because training it how to suffer will only further hurt my job prospects. Nice try, no dice.
Is this the beginning of the apps for everything era and finally the SaaS for your LLM begins? Initially we had internet but value came when instead of installed apps, webapps arrived to become SaaS. Now if LLMs can use specific remote MCP which is another SaaS for your LLM, the remote MCP powered service can charge a subscription to do wonderful things and voila! Let the new golden age of SaaS for LLMs begin and the old fad(replace job XYZ with AI) die already.
I'm more excited I can run now a custom site, hook an MCP for it, and have all the cool intelligence I had to pay for SaaS without having to integrate to them plus govern my data, it's a massive win.
I just see AI assistant coding replicating current SaaS services that I can run internally. If my shop was a specific stack, I could aim to have all my supporting apps in that specific stack using AI assistant coding, simplifying operations, and being able to hook up MCP's to get intelligence from all of them.
Truly, OSS should be more interesting in the next decade for this alone.
We should all thank the chinese companies for releasing so many incredible open weight models. I hope they keep doing it, I dont want to rely on OpenAI, Anthropic or Google for all my future computer interactions.
On one hand, yes this is very cool for a whole host of personal uses. On the other hand giving any company this level of access to as many different personal data sources as are out there scares the shit out of me.
I’d feel a lot better if we had something resembling a comprehensive data privacy law in the United States because I don’t want it to basically be the Wild West for anyone handling whatever personal info doesn’t get covered under HIPAA.
Absolutely agreed, but just wanted to mention that it's essentially the same level of access you would give to Zapier, which is one of their top examples of MCP integrations.
It took many years for online tracking, iframes, sticky cookies and cambridge analytics before things like GDPR came into existence. We have to similarly wait a few years before similar major leaks happen through LLM pipelines/integrations. Sadly, that is the reality we live with.
I'd love a _tip jar_ MCP, where the LLM vendor can automatically tip my website for using its content/feature/service in a query's response. Even if the amount is absolutely minuscule, in aggregate, this might make up for ad revenue losses.
It's perfect, nobody will have time to care about how many 9s your service has because the nondeterministic failure mode now sitting slap-bang in the middle is their problem!
> Now if LLMs can use specific remote MCP which is another SaaS for your LLM, the remote MCP powered service can charge a subscription to do wonderful things and voila!
I've always worked under the assumption the best employees make themselves replaceable via well defined processes and high quality documentation. I have such a hard time understanding why there's so much willingness to integrate irreplaceable SaaS solutions into business processes.
I haven't used AI a ton, but everything I've done has focused on owning my own context, config, etc.. How much are people going to be willing to pay if someone else owns 10+ years of their AI context?
Am I crazy or is owning the context massively valuable?
Hello fellow context owner. I like my modules with their context.sh at their root level. If crafted with care, magic happens. Reciprocally, when AI derails, it's most often due to bad context management and fixed by improving it.
Is each Claude instance a separate individual or is a shared AI? Because I'm not sure I would want an AI that learned about my confidential business information sharing that with anyone else, without my express permission.
This does not sound like it would be learning general information helpful across an industry, but specific, actionable information.
If not available now, is that something that AI vendors are working toward? If so, what is to keep them from using that knowledge to benefit themselves or others of their choosing, rather than the people they are learning from?
While people understand ethics, morals and legality (and ignore them), that does not seem like something that an AI understands in a way that might give them pause before doing an action.
I'm curious what kind of research people are doing that takes 45 minutes of LLM time. Is this a poke at the McKinsey consultant domain?
Perhaps I am just frivolous with my own time, but I tend to use LLMs in a more iterative way for research. I get partial answers, probe for more information, direct the attention of the LLM away from areas I am familiar and towards areas I am less familiar. I feel if I just let it loose for 45 minutes it would spend too much time on areas I do not find valuable.
This seems more like a play for "replacement" rather than "augmentation". Although, I suppose if I had infinite wealth, I could kick of 10+ research agents each taking 45 minutes and then review their output as it became available, then kick off round 2, etc. That is, I could do my process but instead of interactively I could do it asynchronously.
That iterative research process is exactly how I use Google Deep Research since it has a 20/day rate limit. Research a problem, notice some off hand assumption or remark the report made, and fire off another research run asking about it. It depends on what you work on; in my use case I often have to do hours of research for 30 minutes of work like when integrating a bunch of different vendors’ APIs or pouring over datasheets for EE, so it’s worth firing off research and then working on something else for 10-20 minutes (it helps that the Gemini app fires off a push notification when the report is done - Anthropic please do this! Even for requests made from the web app).
As for long research times, one thing I’ve been using it for is historical research on old books. Gemini DeepResearch was the first one able to properly explain the nuances of identifying a chimeral first edition Origin of Species after taking half an hour and reading 400 sources. It went into all the important details like spelling errors and the properties of chimeral FY2** copies found in various libraries around the world.
"To start, you can choose from Integrations for 10 popular services, including Atlassian’s Jira and Confluence, Zapier, Cloudflare, Intercom, Asana, Square, Sentry, PayPal, Linear, and Plaid. ... Each integration drastically expands what Claude can do."
Give us an LLM with better reasoning capabilities, please! All this other stuff just feels like a distraction.
Building integrations is a more predictable way of developing a smaller competitive advantage versus research. I think most of the leading AI companies are adopting a multi-arm strategy of research + product/ecosystem development to balance their risks.
I disagree. They can walk and chew gum, do both things at once. And this practical stuff is very important.
I've been using the Atlassian MCP for nearly a month now, and it's completely changed (and eliminated) the feeling of having an overwhelming backlog.
I can have it do things like "find all the tickets related to profile editing and combine them into one epic" where it works perfectly. Or "help me prioritize the 15 tickets assigned to me this sprint" and it'll actually go through and suggest "maybe you can do these two tickets first since they seem smaller, then do this big one" – i haven't hooked it up to my calendar yet.
But I'd love for it to suggest things like "do this one ticket that requires a lot of heads down time on wednesday since you don't have any meetings. I can create a block on your calendar so that nobody will schedule a meeting then"
Those are all superhuman things that can be done with MCP and a smart model.
I've defined rules in cursor that say "when I ask you to mark something ready for test, change the status and assign it to <x person>, and leave a comment summarizing the changes"
If you look at my JIRA comments now, you'd wonder how I had so much time to write such thorough comments. I don't, Cursor and whatever model is doing it for me.
It's been an absolute game changer. MCP is going to be what the App store was to mobile. Yes you can get by without it, but actually hooking into all your daily tool is when this stuff gets insanely valuable in a practical sense.
> If you look at my JIRA comments now, you'd wonder how I had so much time to write such thorough comments. I don't, Cursor and whatever model is doing it for me.
Joking aside, I do believe we are moving into a era where we have LLMs write for each other and humans have a dedicated TL;DR. This includes code with a lot of comments or design styles that might seem obvious or stupid but can help another LLM.
JIRA is more than just ticket management for most big orgs. It provides a reporting interface for business with long-term planning capabilities. A lot of the annoying things that devs have to do in JIRA is often there to make those functions more valuable. In other cases it is a compliance thing as well. Some certifications necessary for enterprise sales require audit trails for all code changes, from the bug report to the code commit. JIRA provides the integration and reporting necessary for that.
Unless you can provide the same visibility, long-term planning features and compliance aspects of JIRA on top of you sqlite db, you won't compete with JIRA. But if you do add those things on top of SQLite and LLMs, you probably have a solid business idea. But you'd first need to understand JIRA well enough to know why they are there in the first place.
Well I had half a mind to not tell them to see what they’d say, but I also was excited to show everyone so they can also be empowered with it.
One of them said “yeah I was wondering cuz you never write that much” - as a leader, I actually don’t set a good example of how to leave quality JIRA comments. And my view with all these things is that I have to lead by example, not by orders.
With the help of these kinds of tools, we can improve the quality of these comments. And I wouldn’t expect others to write them manually, more that I wanted to show that everyone’s use of JIRA on the team can improve.
Notice they commented on the quantity, not the quality?
I don't think it's good leadership to unleash drivel on an organisation, have people waste time reading and perhaps replying to it, thinking it's something important and thoughtful coming from atonse.
Good thing you told them though, now they can ignore it.
It sure seems like the next evolution of Jira though. Designed to waste everyones time, picked by "leaders" that don't use it.
Why not spam tickets with LLM drivel? They are perfect to pick up on all the inconsistency in the PM insanity driven custom designed workflow - and comment on it tagging a bunch of stray people seen in the ticket history, the universal exit hatch.
In another comment I mentioned that I ask for it to be concise.
Also, a lot of the kinds of comments are things like, when you combine a bunch of tickets, leaving comments on the cancelled tickets to show why they were cancelled.
Someone please shoot me if my PM ever gets this idea in his head of using LLM slop to spam tickets with en masse.
There's nothing I hate more than people sending me their AI messages, be it in a ticket or a PR or even on Slack. I'm forced to engage and spend effort on something it took them all of 3 seconds to generate without even proofreading what they're sending me says. The amount of times I've had to ask 11 clarifying questions because their message has 11 contradictions within itself is maddening to the highest degree.
The worst is when I call out one of these numerous contradictions, and the reply is "oh haha, stupid Claude :)", makes my blood boil and at the same time amazes me that someone has so little pride and respect for their fellow humans to do crap like that.
Sounds like your coworkers might be abusing things here.
I’m not remotely interested in throwing random slop in there.
In fact, we did try a year ago to have AI help write our tickets and it was very clear that they were AI generated. There was way too much nonsense in there that wasn’t relevant to our product.
Honestly, that backlog management idea is probably the first time an MCP actually sounded appealing to me.
I'm not in that world at the moment, but I've been the lead on several projects where the backlog has became a dumping ground of years of neglect. You end up with this tiered backlog thing where one level of backlog gets too big so you create a second tier of backlog for the stuff you are actually going to work on. Pretty soon you end up with duplicates in the second tier backlog for items already in the base level backlog since no one even looks at that old backlog anymore.
I've done a lot of tidy up myself when I inherit this kind of mess, just closing tickets we definitely will never get to, de-duping, adding context when available, grouping into epics, tagging with relevant "tech-debt", "security", "bug", "automation", etc. But when there are 100s of tickets it is a slog. Having an LLM do this makes so much sense.
I have Claude hooked up to our project management system, GitHub, and my calendar (among other things). It's already proving extremely useful for various project management tasks.
Lots of reported security issues with MCP servers seemed to be mitigated by their local-only setup. These MCP implementations are remotely accessible, do they address security differently?
Largely, yes -- one of the big issues with using other people's random MCP servers is that they are run by default as a system process, even if they only need to speak over an API. Remote MCP mitigates this by not running any untrusted code locally.
What it _doesn't_ seem to yet mitigate is prompt injection attacks, where a tool call description of one tool convinces the model to do something it shouldn't (like send sensitive data to a server owned by the attacker.) I think these concerns are a little bit overblown though; things like pypi and the Chrome Extension store scare me more and it doesn't stop them from mostly working.
They offhand mention OAuth integration in their discussion of Cloudflare integrated solutions. I can't see how that would be any less secure than any other OAuth protected API offering.
So any chat to Claude will now just auto-activate web search to be included? What if I try to use it just as a search engine exclusively? Also will proxies like Openrouter have access to the web search capabilities?
265 comments
[ 4.1 ms ] story [ 208 ms ] threadLots of people are making moves in this space (including Anthropic), but nothing has broken through to the mainstream.
Why can't one set up a prompt, test it against a file, then once it is working, apply it to each file in a folder in a batch process which then provides the output as a single collective file?
Not sure what OS you're on, but in Windows it might look like this:
FOR %%F IN (*.txt) DO (TYPE "%%F" | llm -s "execute this prompt" >> "output.txt)
FOR %%F IN (*.pdf) DO (llm -a %%F -s "execute this prompt" >> output.txt)
The limits are still buggy responses - Claude often gets stuck in a useless loop if you overfeed it with files - and lack of consistency. Sometimes hand-holding is needed to get the result you want. And it's slow.
But when it works it's amazing. If the issues and limitations were solved, this would be a complete game changer.
We're starting to get somewhat self-generating automation and complex agenting, with access to all of the world's public APIs and search resources, controlled by natural language.
I can't see the edges of what could be possible with this. It's limited and clunky for now, but the potential is astonishing - at least as radical an invention as the web was.
I also use this method for doing code prototyping by giving it the path to files in the local working copy of my repo. Really cool to see it make changes in a vite project and it just hot reloads. Then I make tweaks or commit changes as usual.
LLM-desktop interfaces make great demos, but they are too slow to be usable in practice.
Tackling individual use-cases is supposed to be something for third party "ecosystem" companies to go after, not the mothership itself.
Not sure what Anthropic and co can do about that, but integrations feel like a step in the wrong direction. Whenever I've tried tool use, it was orders of magnitude more expensive and generally inferior to a simple model call with curated context from SerpApi and such.
I bet there are better / less arcane tools, but I think powerful and fast mechanisms for managing context are key and for me, that's really just powerful text editing features.
A truly useful AI assistant has context on my last 100,000 emails - and also recalls the details of each individual one perfectly, without confusion or hallucination.
Obviously I’m setting a high bar here; I guess what I’m saying is “yes, and”
Throw in all context --> ask it what is important for problem XYZ --> curate what it tells you, and feed that to another model to actually solve XYZ
Different to what this integration is pushing, the LLMs usage in production based products where high accuracy is a requirement (99%), you have to give a very limited tool set to get any degree of success.
I’m a bit skeptical that it’s gonna work out of the box because of the amount of custom fields that seem to be involved to make successful API requests in our case.
But I would welcome, not having to solve this problem. Jira’s interface is among the worst of all the ticket tracking applications I have encountered.
But, I have found using a LM conversation paired within enough context about what is involved for successful POSTs against the API allow me to create update and relate issues via curl.
It’s begging for a chat based LLM solution like this. I’d just prefer the underlying model not be locked to a vendor.
Atlassian should be solving this for its customers.
I hadn't realized but the new integration seems to actually just be an official, closed-source MCP produced *by* Atlassian.
sooperset's MCP is MIT licensed, so I wonder how much of the Atlassian edition is just a lift of that.
There's a comment [1] on the actual integration page asking about custom fields, which I think is possibly a big issue.
At first I thought the open-source version would get crushed by an actual Atlassian release, but not if Atlassian doesn't offer all the support for it to work really well no matter what customizations are fitted into each instance.
My hypothesis is that it takes custom code to make this work, and using the off-the-shelf for Jira won't work. Hoping to be proven wrong though, as it would be less work for me on that front.
[1] https://community.atlassian.com/forums/Atlassian-Platform-ar...
That or they're pulling an OpenAI and launching a feature that isn't actually fully live.
> in beta on the Max, Team, and Enterprise plans, and will soon be available on Pro
LLMs were always a fun novelty for me until OpenAI DeepResearch which started to actually come up with useful results on more complex programming questions (where I needed to write all the code by hand but had to pull together lots of different libraries and APIs), but it was limited to 10/month for the cheaper plan. Then Google Deep Research upgraded to 2.5 Pro and with paid usage limits of 20/day, which allowed me to just throw everything at it to the point where I'm still working through reports that are a week or more old. Oh and it searched up to 400 sources at a time, significantly more than OpenAI which made it quite useful in historical research like identifying first edition copies of books.
Now Claude is releasing the same research feature with integrations (excited to check out the Cloudflare MCP auth solution and hoping Val.town gets something similar), and a run time of up to 45 minutes. The pace of change was overwhelming half a year ago, now it's just getting ridiculous.
However, unfortunately, I cannot shower much praise on Claude 3.7. And if you (or anyone) asks why - 3.7 seems much better than 3.5, surely? - Then I’m moderately sure that you use Claude much more for coding than for any kind of conversation. In my opinion, even 3.5 Haiku (which is available for free during high loads) is better than 3.7 Sonnet.
Here’s a simple test. Try asking 3.7 to intuitively explain anything technical - say, mass dominated vs spring dominated oscillations. I’m a mechanical engineer who studied this stuff and I could not understand 3.7’s analogies.
I understand that coders are the largest single group of Claude’s users, but Claude went from being my most used app to being used only after both chatgpt and Gemini, something that I absolutely regret.
I haven't been able to use Claude research yet (it's not rolled out to the Pro tier) but o1 -> o3 deep research was a massive jump IMHO. It still isn't perfect but o1 would often give me trash results but o3 deep research actually starts to be useful.
3.5->3.7 (even with extended thinking) felt like a nothingburger.
Like, ask it a simple question and it comes up with a full repo, complete with a README and a Makefile, when all you wanted to know was how efficient a particular algorithm would be in the included code.
Can't wait until the add research to the Pro plan because, you know, I have questions...
If you pay for a subscription then they don’t have an incentive to use more tokens for the same answer.
It’s definitely because feedback from people has “taught” it that more boilerplate is better. It’s the same reason ChatGPT is annoyingly complementary.
I prefer Gemini 2.5 pro for all code now
But Gemini is at least as overactive as Claude, sometimes even more overactive when it comes to something like comment spam.
Of course, this can be fixed with prompting. And sometimes it feels sheepish complaining about the machine god doing most of my chore work that didn't even exist a couple years ago.
That's not really true, since your prompts are also getting better. Better input leads to better output remains true, even with LLMs (when you see it as a tool).
However in this specific example, I don't remember if it was chatgpt or gemini or 3.5 Haiku but the other(s) explained it well enough. I think I re-asked 3.5 Haiku at a later point of time, and to my complete non-surprise, it gave an answer that was quite decent.
1 - For example, the field of DIY audio - which was funnily enough the source of my question. I'm no speaker designer, but combining creativity with engineering basics/rules of thumb seems to be something LLms struggle with terribly. Ask them to design a speaker and they come up with the most vanilla, tired, textbook design - despite several existing market products that are already so much ahead/innovative.
I'm confident that if you asked an LLM an identical question for which there is more discourse - eg make an interesting/innovative phone - you'd get relatively much better results.
I am a novice, maybe that's why I liked it.
For example, I’m not a domain expert, but I was looking for an RC motor for a toy project and OpenAI had happily tried to source a few, with Deep Research. Only the best candidate it had picked contained an obvious typo in the motor spec (68 grams instead of 680 grams), which is just impossible for a motor of specified dimensions.
It's funny you say that because I was going to echo your parents sentiment and point out it's exactly the same with any news article you read.
The majority if content these LLMs are consuming is not from domain experts.
This isn’t just “shit in, shit out”. Hallucination is real and still problematic.
That said, the “AI all the things” mandates could be the lever that ultimately accomplishes what 100+ PjMs couldn’t - making people write issues as if they really mattered. Because garbage in, garbage out.
If it was actually good they would've named it 4.0, the fact that they went from 3.5 to 3.7 (weird jump) speaks volumes imo.
3.7 is definitively better at coding but you feel it lost a bit of maneuverability at other domains. For someone who wants code generated, it doesn't matter but I've found myself using DeepSeek first and then getting code output by 3.7.
Deep Research hasn't really been that good for me. Maybe I'm just using it wrong?
Example: I want the precipitation in mm and monthly high and low temperature in C for the top 250 most populous cities in North America.
To me, this prompt seems like a pretty anodyne and obvious task for Deep Research. It's long, tedious, but mostly coming from well structured data sources (wikipedia) across two languages at most.
But when I put this in to any of the various models, I mostly get back ways to go and find that data myself. Like, I know how to look at Wikipedia, it's that I don't want to comb through 250 pages manually or try to write a script to handle all the HTML boxes. I want the LLM/model to do this days long tedious task for me.
My perspective on this is that if Deep Research can't do something, you should do it yourself and put the results on the internet. It'll help other humans and AIs trying to do the same task.
The project requires the full list of every known city in the western hemisphere and also Japan, Korea, and Taiwan. But that dataset is just maddeningly large, if it is possible at all. Like, I expect it to take me years, as I have to do a lot of translations. So, I figured that I'd be nice and just as for the top 250 for the various models.
There's a lot more data that we're trying to get too and I'm hoping that I can get approval to post it as its a work thing.
How do you validate its results in that scenario? Just take its word for it?
I'd say that what you're saying is 'synthesis'. The 'Intro/Discussion' sections of a journal article.
For me, 'research' means the work of going through and getting all the data in the first place. Like, going out and collecting dino bones in the hot sun, measuring all the soil samples, etc. - that is research. For me, asking these models to go collate some webpages, I mean, you spend the first weeks of a summer undergrad's time to go do this kid of thing to get them used to the file systems and spruce up their organization skills, see where they are at. Writing the paper up, that's part of research sure, but not the hard part that really matters.
Maybe in a year, they’ll hit the graduate level. But we’re not near PhD level yet
I use it a lot when documentation is vague or outdated. When Gemini/o3 can't figure something out after 2 tries. When I am working with a service/API/framework/whatever that I am very unfamiliar with and I don't even know what to Google search.
I recently asked Chrome to show me how to apply the Knuth-Bendix completion procedure to propositional logic, and I had already formed my own thoughts about how to proceed (I'm building a rewrite system that does automated reasoning).
The response convinced me that I'm not a total idiot.
I'm not an academic and I'm often wrong about theory so the validation is really useful to me.
It is literally stagnated for a year now
All that changed is they connect more apis.
And add a thinking loop with same model powering it
This is the reason it seems fast - nothing really happens except easy things
That being said, isn’t it strange how the community has polar opposite views about this? Did anything like this ever happen before?
Like I wanted to scope how to build a home made TrueNAS Scale unit, it helped me with a avoiding pitfalls like knowing that I needed two GPUs minimum to run the OS and local llms, and speed up config for a CLI back up of my Dropbox locally(it told me to use the right filesystem format over ZFS to make Dropbox client work).
It has researched how I can structure my web app for building payment system on the web(something I knew nothing about) to writing small tools to talk to my document collection and index them into collections in Anki in one day.
All those talks about AI replacing people seemed a little far fetched in 2024. But in 2025, I really think models are getting good enough
If there was truly any innovation still happening in OpenAI, Anthropic, etc., they would be working on models only, not on side features that someone could already develop over a weekend.
Now I'm in a new team where 99% of our oncall tickets come from automated alarms and 80% of them are a subset of a few issues where the root-cause isn't easy to address but there is either nothing to actually do once investigated, or the fix is a one time process that is annoying to run, so the username isn't accurate anymore :)
I still like the change of pace though, 0 worries about sprint tasks or anything else for a week every few months.
Is there a youtube video of ppl using this on complex open source projects like linux kernel or maybe something like pytorch.
How come none of the oss pojects( atleast not the ones i follow) are progressing fast(er) from AI like 'deepresearch'
Hope one day it will be practical to do nightly finetunes of a model per company with all core corporate data stores.
This could create a seamless native model experience that knows about (almost) everything you’re doing.
I'll leave it to you to guess which one is harder to do.
There are now some light versions of fine tuning that don’t update all the model weights but train a small adapter layer called Lora which is way more viable commercially atm in my opinion.
Some of the issues still exist, of course:
* Finetuning takes time and compute; for one-off queries using in-context learning is vastly more efficient (i.e., look it up with RAG).
* Early results with finetuning had trouble reliably memorizing information. We've got a much better idea of how to add information to a model now, though it takes more training data.
* Full finetuning is very VRAM intensive; optimizations like LoRA were initially good at transferring style and not content. Today, LoRA content training is viable but requires training code that supports it [1].
* If you need a very specific memorized result and it's costly to get it wrong, good RAG is pretty much always going to be more efficient, since it injects the exact text in context. (Bad RAG makes the problem worse, of course).
* Finetuning requires more technical knowledge: you've got to understand the hyperparameters, avoid underfitting and overfitting, evaluate the results, etc.
* Finetuning requires more data. RAG works with a handful datapoints; finetuning requires at least three orders of magnitude more data.
* Finetuning requires extra effort to avoid forgetting what the model already knows.
* RAG works pretty well when the task that you are trying to perform is well-represented in the training data.
* RAG works when you don't have direct control over the model (i.e., API use).
* You can't finetune most of the closed models.
* Big, general models have outperformed specialized models over the past couple of years; if it doesn't work now, just wait for OpenAI to make their next model better on your particular task.
On the other hand:
* Finetuning generalizes better.
* Finetuning has more influence on token distribution.
* Finetuning is better at learning new tasks that aren't as present in the pretraining data.
* Finetuning can change the style of output (e.g., instruction training).
* When finetuning pays off, it gives you a bigger moat (no one else has that particular model).
* You control which tasks you are optimizing for, without having to wait for other companies to maybe fix your problems for you.
* You can run a much smaller, faster specialized model because it's been optimized for your tasks.
* Finetuning + RAG outperforms just RAG. Not by a lot, admittedly, but there's some advantages.
Plus the RL Training for reasoning has been demonstrating unexpectedly effective improvements on relatively small amounts of data & compute.
So there's reasons to do both, but the larger investment that finetuning requires means that RAG has generally been more popular. In general, the past couple of years have been won by the bigger models scaling fast, but with finetuning difficulty dropping there is a bit more reason to do your own finetuning.
That said, for the moment the expertise + expense + time of finetuning makes it a tough business proposition if you don't have a very well-defined task to perform, a large dataset to leverage, or other way to get an advantage over the multi-billion dollar investment in the big models.
[1] https://unsloth.ai/blog/contpretraining
1. If you have a large corpus of valuable data not available to the corporations, you can benefit from fine tuning using this data.
2. Otherwise just use RAG.
Fine-tuning makes sense when you need behavioral shifts (style, tone, bias) or are training on data unavailable at runtime.
RAG excels when you want factual augmentation without retraining the whole damn brain.
It's not either/or — it's about cost, latency, use case, and update cycles. But hey, binaries are easier to pitch on a slide.
I had no idea that fine tuning for adding information is viable now. Last I checked (year+ back) it seemed to not work well.
it depends on your data access pattern. If some text goes through LLM input many times, it is more efficient for LLM to be finetuned on it once.
The budget question comes into play as well. Even if text is repetitively fed to the LLM, that might happen over a long enough time compared to finetuning which is a sort of capex that it is financially more accessible.
Now bear in mind, I'm a big proponent of finetuning where applicable and I try to raise awareness to the possibilities it opens. But one cannot deny RAG is a lot more accessible to teams which are likely developers / AI engineers compared to ML engineers/researchers.
It looks like major vendors provide simple API for fine-tuning, so you don't need ML engineers/researchers: https://platform.openai.com/docs/guides/fine-tuning
Setting RAG infra is likely more complicated than that.
Results with this method are significantly more limited compared to all the power open-weight finetuning gives you (and the skillset needed in return).
And in either case don’t forget alignment and evals.
I am not sure I understand why you are so certain that finetuned top market models, built by top researchers will be significantly worse than whatever open source model you pick.
How many epochs do you run?
I worked at a startup where the CEO swore up and down that real-time fine-tuning was the future — that models would continuously update with company data. It sounded cool until you remember: That’s not how LLMs work. It’s not efficient. It’s not flexible. And it’s not even necessary — we already have RAG.
Pipedreams make good pitch decks. But they break when you hit production.
It's a fucking pipedream, this. That's not how LLMs work, it's not efficient, it's not useful (we have RAG for reference augmentation), and it’s not even desirable unless you want your model overfitting on stale, internal narratives every night.
In case the above link doesn't work later on, the page for this demo day is here: https://demo-day.mcp.cloudflare.com/
Because MCP isn’t an API it’s the protocol that defines how the LLM even calls the API in the first place. Without it, all you've got is a chat interface.
A lot of people misunderstand what is the role of MCP. It’s the signaling the LLM uses to reach out of its context window and do things.
Truly, OSS should be more interesting in the next decade for this alone.
I’d feel a lot better if we had something resembling a comprehensive data privacy law in the United States because I don’t want it to basically be the Wild West for anyone handling whatever personal info doesn’t get covered under HIPAA.
It is really cool to witness the velocity of MCP adoption.
I've always worked under the assumption the best employees make themselves replaceable via well defined processes and high quality documentation. I have such a hard time understanding why there's so much willingness to integrate irreplaceable SaaS solutions into business processes.
I haven't used AI a ton, but everything I've done has focused on owning my own context, config, etc.. How much are people going to be willing to pay if someone else owns 10+ years of their AI context?
Am I crazy or is owning the context massively valuable?
This does not sound like it would be learning general information helpful across an industry, but specific, actionable information.
If not available now, is that something that AI vendors are working toward? If so, what is to keep them from using that knowledge to benefit themselves or others of their choosing, rather than the people they are learning from?
While people understand ethics, morals and legality (and ignore them), that does not seem like something that an AI understands in a way that might give them pause before doing an action.
Perhaps I am just frivolous with my own time, but I tend to use LLMs in a more iterative way for research. I get partial answers, probe for more information, direct the attention of the LLM away from areas I am familiar and towards areas I am less familiar. I feel if I just let it loose for 45 minutes it would spend too much time on areas I do not find valuable.
This seems more like a play for "replacement" rather than "augmentation". Although, I suppose if I had infinite wealth, I could kick of 10+ research agents each taking 45 minutes and then review their output as it became available, then kick off round 2, etc. That is, I could do my process but instead of interactively I could do it asynchronously.
As for long research times, one thing I’ve been using it for is historical research on old books. Gemini DeepResearch was the first one able to properly explain the nuances of identifying a chimeral first edition Origin of Species after taking half an hour and reading 400 sources. It went into all the important details like spelling errors and the properties of chimeral FY2** copies found in various libraries around the world.
Give us an LLM with better reasoning capabilities, please! All this other stuff just feels like a distraction.
I've been using the Atlassian MCP for nearly a month now, and it's completely changed (and eliminated) the feeling of having an overwhelming backlog.
I can have it do things like "find all the tickets related to profile editing and combine them into one epic" where it works perfectly. Or "help me prioritize the 15 tickets assigned to me this sprint" and it'll actually go through and suggest "maybe you can do these two tickets first since they seem smaller, then do this big one" – i haven't hooked it up to my calendar yet.
But I'd love for it to suggest things like "do this one ticket that requires a lot of heads down time on wednesday since you don't have any meetings. I can create a block on your calendar so that nobody will schedule a meeting then"
Those are all superhuman things that can be done with MCP and a smart model.
I've defined rules in cursor that say "when I ask you to mark something ready for test, change the status and assign it to <x person>, and leave a comment summarizing the changes"
If you look at my JIRA comments now, you'd wonder how I had so much time to write such thorough comments. I don't, Cursor and whatever model is doing it for me.
It's been an absolute game changer. MCP is going to be what the App store was to mobile. Yes you can get by without it, but actually hooking into all your daily tool is when this stuff gets insanely valuable in a practical sense.
How do your colleagues feel about it?
I also don’t want to read too many unnecessary words.
Can’t we point an LLM to a sqlite db and tell it to treat it as an issue tracking db and have everyone do the same.
The service (jira) would materialize inside the LLMs then.
Why even use abstractions like tickets etc. Ask LLM what to do.
Unless you can provide the same visibility, long-term planning features and compliance aspects of JIRA on top of you sqlite db, you won't compete with JIRA. But if you do add those things on top of SQLite and LLMs, you probably have a solid business idea. But you'd first need to understand JIRA well enough to know why they are there in the first place.
[0] https://en.wikipedia.org/w/index.php?title=Wikipedia:FENCE
One of them said “yeah I was wondering cuz you never write that much” - as a leader, I actually don’t set a good example of how to leave quality JIRA comments. And my view with all these things is that I have to lead by example, not by orders.
With the help of these kinds of tools, we can improve the quality of these comments. And I wouldn’t expect others to write them manually, more that I wanted to show that everyone’s use of JIRA on the team can improve.
I don't think it's good leadership to unleash drivel on an organisation, have people waste time reading and perhaps replying to it, thinking it's something important and thoughtful coming from atonse.
Good thing you told them though, now they can ignore it.
Also, a lot of the kinds of comments are things like, when you combine a bunch of tickets, leaving comments on the cancelled tickets to show why they were cancelled.
In the past, that info simply wouldn’t be there.
There's nothing I hate more than people sending me their AI messages, be it in a ticket or a PR or even on Slack. I'm forced to engage and spend effort on something it took them all of 3 seconds to generate without even proofreading what they're sending me says. The amount of times I've had to ask 11 clarifying questions because their message has 11 contradictions within itself is maddening to the highest degree.
The worst is when I call out one of these numerous contradictions, and the reply is "oh haha, stupid Claude :)", makes my blood boil and at the same time amazes me that someone has so little pride and respect for their fellow humans to do crap like that.
I’m not remotely interested in throwing random slop in there.
In fact, we did try a year ago to have AI help write our tickets and it was very clear that they were AI generated. There was way too much nonsense in there that wasn’t relevant to our product.
So we don’t do that.
I'm not in that world at the moment, but I've been the lead on several projects where the backlog has became a dumping ground of years of neglect. You end up with this tiered backlog thing where one level of backlog gets too big so you create a second tier of backlog for the stuff you are actually going to work on. Pretty soon you end up with duplicates in the second tier backlog for items already in the base level backlog since no one even looks at that old backlog anymore.
I've done a lot of tidy up myself when I inherit this kind of mess, just closing tickets we definitely will never get to, de-duping, adding context when available, grouping into epics, tagging with relevant "tech-debt", "security", "bug", "automation", etc. But when there are 100s of tickets it is a slog. Having an LLM do this makes so much sense.
What it _doesn't_ seem to yet mitigate is prompt injection attacks, where a tool call description of one tool convinces the model to do something it shouldn't (like send sensitive data to a server owned by the attacker.) I think these concerns are a little bit overblown though; things like pypi and the Chrome Extension store scare me more and it doesn't stop them from mostly working.
I love MCP (it’s way better than plain Claude) but even that runs into context walls.