We'd never hire someone who just makes stuff up (or at least keep them employed for long). Why are we okay with calling "AI" tools like this anything other than curious research projects?

Can't we just send LLMs back to the drawing board until they have some semblance of reliability?

A report full of factual errors that a careful intern wouldn't make is worse than useless (yes, yes, I've mentored interns).

If the hard part is the language, then do the research yourself, write an outline, and have the LLM turn it into complete sentences. That would at least be faster.

Here's the thing, though: If you do that, you're effectively proving that prose style is the low-value part of the work, and may be unnecessary. Which, as much as it pains to me say as a former English major, is largely true.

This is why I don't use AI for anything that requires a "correct" answer. I use it to re-write paragraphs or sentences to improve readability etc, but I stop short of trusting any piece of info that comes out from AI.

I wonder if it’s carried over too much of that ‘helpful’ DNA from 4o’s RLHF. In that case, maybe asking for 500 words was the difficult part — it just didn’t have enough to say based on one SO post and one article, but the overall directives assume there is, and so the model is put into a place where it must publish..

Put another way, it seems this model faithfully replicates the incentives most academics have — publish a positive result, or get dinged. :)

Did it pick up your HN comments? Kadua claims that’s more than enough to roast me, … and it’s not wrong. It seems like there’s enough detail about you (or me) there to do a better job summarizing.

Artificial dementia...

Some parties are releasing products much earlier than the ability to ship well working products (I am not sure that their legal cover will be so solid), but database aided outputs should and could become a strong limit to that phenomenon of remembering badly. Very linearly, like humans: get an idea, then compare it to the data - it is due diligence and part of the verification process in reasoning. It is as if some moves outside linear pure product progress reasoning are swaying the RnD towards directions outside the primary concerns. It's a form of procrastination.

That would be exactly my verdict of any product based on LLMs in the past few years.

You might find it amusing to compare it to: https://hn-wrapped.kadoa.com/timabdulla

(Ref:https://news.ycombinator.com/item?id=42857604)

Still useful for the odd task here and there, but not as useful as all the money being invested in this (except for the companies getting that money, that is).

edit: actual example of something I'd expect a real AI to be able to solve by itself, but currently LLMs fail miserably https://x.com/RadishHarmers/status/1885884032220643587

This is why I haven't found AI tools very useful. I find my self spending more time verifying and fixing it's answers than I would have just doing or learning the darn thing myself.

I always thought there was a kind of NP-flavor to the problems for which LLMs-like AI are helpful in practice, in the sense that solving the problem may be hard but checking the solution must be fast.

Unless the domain can accomodate errors/hallucination, checking the solution (by a human) should be exponentially faster than finding it (by some AI) otherwise there's little practical gain.

A year back that number was 30%, and a couple of years back it was 60%. There will be a point where it'll be good enough. There are also better and better ways to verify answers these days.

It'll never be a solution for everything, but that's similar to many engineering problems we have: for example, ORMs aren't great for all types of queries, but they're sufficient for a good part of them.

I'd say they don't know anything.

An LLM base model, before it is post-trained with RL, just has access to a sliced and diced corpus of human output. Take the contents of 4chan and WikiPedia, put in blender and mix and chop into "training sample" sized bites, then learn the statistical regularities of this blended mess. It is what it is - not exactly what I'd call a knowledge base, even though there are bits of knowledge in there.

When you add RL-based post-training for reasoning, all you are doing is trying get the model to be more selective when you are sampling from it - encouraging it to suppress some statistics, and emphazise others, such that when you sample from it the output looks more like valid reasoning steps and/or conclusions, per the verified reasoning examples you train it on.

I'm well aware of how useful RL-tuned models (whatever the goal) can be, but at the end of the day all they are doing is taking a statistical babbler and saying "try to output patterns more like this". It's not exactly a recipe for factuality or rationality - we've just gone from hallucination-prone base models, to gaslighting-prone RL-tuned "reasoning" models that output stuff that sounds like coherent reasoning.

What missing from all of this - what makes it different from how animals learn - it that the model has no experience of it's own, no autonomy or motivation to explore, learn and verify, and hence no episodic memories of how it learnt something (tried it and ran controlled experiments, or just overheard it on the bus), and what that implies about it's trustworthiness.

It's amazing that LLMs work as well as they do, a reflection of how much of what we do can be accomplished by reactive pattern matching, but if you want to go beyond that to something that can learn and discern the truth for itself, this seems the wrong paradigm altogether.

"Limitations Deep research unlocks significant new capabilities, but it’s still early and has limitations. It can sometimes hallucinate facts in responses or make incorrect inferences"

How do I know which parts are false? It will take as long to verify as to research!

More importantly, I think, is that they are incapable of not doing so. Have we figured out how to make an LLM realize and answer that it doesn't know an answer?

"PS: The name Triton was coined in mid-2019 when I released my PhD paper on the subject. I chose not to rename the project when the "TensorRT Inference Server" was rebranded as "Triton Inference Server" a year later since it's the only thing that ties my helpful PhD advisors to the project."

Would have loved to try it when they released it, but I'm apparently in the wrong country. I think it's not available outside the US (?). OpenAI and DeepSeek have no such issues. It's a bummer really, I'm happy paying for this but they don't want me to.

I think you have to prefix the query with "@Meta AI", hope this helps

They seem to be ok with outsourcing any and all creativity to a language model, so it’s not surprising that they can’t come up with unique names themselves.

We found in our experience it is pretty hard to force LLM to do something in proper depth, and OAI's deep research definitely feels like one of the first examples from big labs on how this can be done. What we typically see is that it is not even the "agent" part that is hard to do, but how to force model to not "forget" to go deep...

The majority of human written consultant reports are already complete rubbish. Low accuracy, low signal-to-noise, generic platitudes in a quantity-over-quality format.

LLMs are innoculating people to this kind of low information value content.

People who produce LLM quality output, are now being accused of using LLMs, and can no longer pretend to be adding value.

The result of this is going to be higher quality expectations from consultants and a shaking out of people who produce word vommit rather than accurate, insightful, contextually relevent information.

Oh Sweet summer child.

It's the only tool/system (I won't call it an LLM) in their released benchmarks that has access to tools and the web. So, I'd wager the performance gains are strictly due to that.

If an LLM (o3) is too expensive to be released to the public, why would you use it in a tool that has to make hundreds of inference calls to it to answer a single question? You'd use a much cheaper model. Most likely o3-mini or o1-mini combined with o4-mini for some tasks.

What makes you believe that?

Effectiveness in this task environment is well beyond the specific model involved, no? Plus they'd be fools (IMHO) to only use one size of model for each step in a research task -- sure, o3 might be an advantage when synthesizing a final answer or choosing between conflicting sources, but there are many, many steps required to get to that point.

Hopefully helpful, happy to test others for you :)

-- Raw output --

Understood. I will conduct a detailed technical analysis of next-generation particle collider projects, focusing on the Future Circular Collider (FCC), International Linear Collider (ILC), Compact Linear Collider (CLIC), Muon Collider proposals, and other significant initiatives.

The report will include: - Planned energy ranges and collision types - Estimated construction and operation timelines - Technical advantages and challenges - Approximate costs - Key physics goals - Review of technical design reports and feasibility studies - International collaboration and support - Comparative analysis on feasibility, cost-benefit, scientific potential, infrastructure needs, and environmental impact - Long-term considerations, including upgrade paths and integration with existing facilities

The analysis will be formatted as a structured technical summary, suitable for presentation at a topical meeting of particle physicists, with comparative tables and figures where relevant.

I will update you when the report is ready.

# Next-Generation Particle Collider Projects: A Comparative Technical Analysis

## Introduction High-energy physics is charting plans for new colliders to extend beyond the LHC and its High-Luminosity upgrade (HL-LHC). The discovery of the 125 GeV Higgs boson has motivated proposals for a dedicated “Higgs factory” lepton collider to study its properties with precision, as well as future energy-frontier machines to search for new physics. Several next-generation collider initiatives are under consideration internationally as of 2024. These include CERN’s proposed *Future Circular Collider (FCC)*, Japan’s *International Linear Collider (ILC)*, CERN’s *Compact Linear Collider (CLIC)*, various designs for a *Muon Collider*, China’s *Circular Electron-Positron Collider (CEPC)* and its successor *Super Proton-Proton Collider (SppC)*, among others. Each proposal differs in collision type (electron-positron, proton-proton, muon-muon, etc.), energy scale, technology, timeline, cost, and physics focus. This summary reviews each project’s key parameters – *planned energy ranges, collision types, timeline, technical advantages/challenges, cost, and physics goals* – based on technical design reports and feasibility studies. A comparative analysis then contrasts their *technical feasibility, cost-benefit, scientific potential for discoveries, timeline to first data, infrastructure needs, and environmental impact*, highlighting the relative strengths and weaknesses of each approach. We also discuss long-term implications such as upgrade paths, flexibility for future modifications, and integration with existing infrastructure.

(Citations refer to official reports and peer-reviewed sources using the format 【source†lines】.)

## Future Circular Collider (FCC) – CERN - *Type and Energy:* The FCC is a *proposed 100 km circular collider* at CERN that would be realized in stages. The first stage, *FCC-ee*, is an electron-positron ($e^+e^-$) collider with center-of-mass energy tunable from ~90 GeV up to 350–365 GeV, covering the Z boson pole, WW threshold, Higgs production (240 GeV), and top-quark pair threshold (~350 GeV). A second stage, *FCC-hh*, would use the same tunnel for a proton-proton collider at up to *100 TeV* center-of-mass energy (an order of magnitude above the LHC’s 14 TeV). Heavy-ion collisions (e.g. Pb–Pb) are also envisioned. An *FCC-eh* option (electron-hadron collisions) is considered by adding a high-energy electron injector to collide with the proton beam. This integrated FCC program thus spans both *precision lepton* collisions and *energy-frontier hadron* collisions.

- *Timeline:* The conceptual schedule foresees *FCC-ee construction in the 2030s* and a start of operations by around *2040* (as the LHC/HL-LHC program winds down). According to the FCC Conceptual Design Report, an $e^+e^-$ Higgs factory could begin delivering physics in ~2040, running for 15–20 years. The *hadron collider FCC-hh* would be constructed subsequently (using the same tunnel and upgraded infrastructure), aiming for *first proton-proton collisions in the late 2050s】. This staged approach (lepton collider first, hadron later) mirrors the successful *LEP–LHC sequence*, leveraging the $e^+e^-$ machine to produce great precision data (and to build infrastructure) before pushing to the highest energies with the hadron machine. ...

(Too long for HN to write more)

So you wouldn't use this tool for those types of use cases.

But still, a valid point. I recall I once wanted to compare Hydroflask, Klean Kanteen and Thermos to see how they perform for hot/cold drinks. I was looking specifically for articles/posts where people had performed actual measurements. But those were very hard to find, with almost all Google hits being generic comparisons with no hard data. That didn't stop them from ranking ("Hydroflask is better for warm drinks!")

Would I be able to get this to ignore all of those and use only ones where actual experiments were performed. And moreover, filter out duplicates (e.g. one guy does an experiment, and several other bloggers link to his post and repeat his findings in their own posts - it's one experiment but with many search results).

That depends entirely on how you choose to define "AGI".

Moravec detailed the idea of a landscape of human capabilities slowly being submerged by AI capabilities, and the point at which AI can do anything a human can, in practice or in principle, we'll know for certain we've reached truly general AI. This idea includes things like feeling pain and pleasure, planning, complex social, oral, and ethical dynamics, and anything else you can possibly think of as relevant to human intelligence. Deep Research is just another island being slowly submerged by the relentless and relentlessly accelerating flood.

The set of natural transformations between two functors F,G ⁣:C→DF,G:C→D can be expressed as the end Nat(F,G)≅∫AHomD(F(A),G(A)). Nat(F,G)≅∫A HomD (F(A),G(A)).

Define set of natural cotransformations from FF to GG to be the coend CoNat(F,G)≅∫AHomD(F(A),G(A)). CoNat(F,G)≅∫AHomD (F(A),G(A)).

Let: - F=B∙(Σ4)∗/F=B∙ (Σ4 )∗/ be the under ∞∞-category of the nerve of the delooping of the symmetric group Σ4Σ4 on 4 letters under the unique 00-simplex ∗∗ of B∙Σ4B∙ Σ4 . - G=B∙(Σ7)∗/G=B∙ (Σ7 )∗/ be the under ∞∞-category nerve of the delooping of the symmetric group Σ7Σ7 on 7 letters under the unique 00-simplex ∗∗ of B∙Σ7B∙ Σ7 .

How many natural cotransformations are there between FF and GG?

pass rate really only matters in context of the difficulty of the tasks

And eyeballing the benchmarks, it'll probably reach a >50% rate per query by the end of the year. Seems to double every model or two.

I mean I too can complain that my iPhone doesn’t automatically screen out spammers and send my mom flowers on Mother’s Day.

OpenAI is very much in an existential crisis and their poor execution is not helping their cause. Operator or “deep research” should be able to assume the role of a Pro user, run a quick test, and reliably report on whether this is working before the press release right?

I wouldn’t even be surprised if a law is passed requiring sites to provide equal access to humans whether accessed directly or via these models.

It’s too important an innovation to stall, especially considering the US’s competitors (China) won’t respect robots.txt either.

How would you know its a crawler?

They do say "Currently, deep research can access the open web...", so maybe "open" there implies something significant. Like, "websites that have agreements with OpenAI and/or do not enforce norobot policies".

I've used Code Interpreter to compile and run C code - https://simonwillison.net/2024/Mar/23/building-c-extensions-... - and I've managed to get it to run JavaScript (by uploading a Deno binary) and even Lua and PHP in the past as well: https://til.simonwillison.net/llms/code-interpreter-expansio...

These systems serve best at augmenting information discovery. When I'm tackling a new area or looking for the right terminology, these models provide a quick shortcut because they have good probabilistic "understanding" of my naive, jargon-free description. This allows me to pull in all of the jargon for the area of research I'm interested in, and move on to actually useful resources, whether that be journal articles, textbooks, or - rarely - online posts/blogs/videos.

the current "meta" is probably something like Elicit + notebookLM + Claude for accelerating understanding of complex topics and extracting useful parts. But, again, each step requires that I am closely involved, from selecting the "correct" papers, to carefully aggregating and grooming the information pulled in from notebookLM, to judging the the usefulness of Claude's attempts to extract what I have asked for

I thought funding was the biggest barrier to great research

  > combine (a number of things) into a coherent whole: pupils should synthesize the data they have gathered | Darwinian theory has been synthesized with modern genetics.

The first was on a topic with which I am deeply familiar -- myself -- and it made three factual errors in a 500-word report: https://news.ycombinator.com/item?id=42916899

The second was a task to do an industry analysis on a space in which I worked for about ten years. I think its overall synthesis was good (it accorded with my understanding of the space), but there were a number of errors in the statistics and supporting evidence it compiled, based upon my random review of the source material.

I think the product is cool and will definitely be helpful, but I would still recommend verifying its outputs. I think the process of verification is less time-consuming than the process of researching and writing, so that is likely an acceptable compromise in many cases.

> when Google released Gemini to have a product in the space.

Bard preceded Gemini.

I highly doubt that the concerns of internal political commissars were holding up this particular openai release.

That’s probably a feature not a bug (from OpenAI’s perspective…).

Few people care or even know about various advances in various specialized fields. It's enough if AI simply seeps into various applications in boring and non-flashy ways for it to have significant effects that will affect a wider range of people, whether they get hyped by the news announcements or not. Jobs etc.

An analogy: the Internet as such is not very exciting nowadays, certainly not in the way it was exciting in the 90s with all the news segments about surfing the information superhighway or whatever. There was a lot of buzz around the web, but then it got normalized. It didn't disappear, it just got taken for granted. No average person got excited around HTML5 or IPv6. It just chugs along in the background. AI will similarly simply build into the fabric of how things get done. Sometimes visibly to the average person, sometimes just behind the scenes.

I want to understand does this mean it's ethical for anyone to create a research AI tool that will go through arXiv and related GitHub repo and use it to solve problems, implement ideas like cursor.

More seriously, it’s unclear why one should be excited by the prospect of AGI, especially when instrumentalized by corporations and authoritarian governments.

Man, I've got a great deal on some oceanfront property in Wyoming for you.

Maybe you can define what "AGI" really means and what the end-game and the economic implications are when 'AGI" is some-what achieved? OpenAI somehow believes that they haven't achieved "AGI" yet, which they continue to do this on purpose for obvious reasons.

The first hint I will give you is that it certainly won't be a utopia.

I don't feel like OpenAI does a good job of getting me excited either.

Find the perfect snowboard? How can that idea get pitched and make the final cut for a $200 a month service? The NFL kicker example is also completely ridiculous.

The business and UX example seems interesting. Would love to see more.

**with browsing + python tools

Maybe we have different definitions of scaling?

Different pre-cooked prompts and filters don’t really amount to new products either, despite them being marketed as such. It’s like adobe treating each tool in photoshop as its own product.

This is like arguing an Airbus cannot possibly fly because it is 165 tonnes of aluminum, steel and plastic.

The proof is in the fact that it flies, not what it is constructed from.