OpenAI's Deep Research Is Actually Good, But Google's Is...
ฝัง
- เผยแพร่เมื่อ 8 ก.พ. 2025
- I tried testing OpenAI Deep Research and Google Gemini's Deep Research yet again centered around the topic of my PhD work. This time however, a clear winner between the two emerges.
Thanks for continuing to do this Kyle - voices like yours are necessary
Thank you for the support!
For some reason I couldn't stop laughing when Google's researching googled how to do research.
lol I also have laughing to video by Kyle for reasons humourous. I am human laughing - just like you.
counterimmunoelectrophoresis antidisestablishmentarianisms oesophagogastroduodenoscopies
pneumonoultramicroscopicsilicovolcanoconios-is
methylenedioxymetamphetamines
ethylenediamenetetraacetates
pseudopseudohypoparathyroidisms
thyroparathyroidectomized
pyschoneuroendocrinologically
🤣🤣🤣🤣🤣🤣
Can we finally admit that deep research is the first use case in which we can say AI is better than the average human by a landslide? It might not be able to beat somebody at their native field but if you give them tasks for an unrelated field the AI will beat the human every single time
I'm sorry what? As the average human? AI has surpassed humans since a year or so.
Sorry you just woke up. That has been well over a year now
Technically, AI has had use cases where it surpasses humans by a landslide since computers were invented and they could do calculations faster and more reliably than humans.
Deep learning has also had many use cases at destroying humans, just look at how it beat the world champion at Go or how it’s able to determine the 3d structures of proteins.
Maybe this is the first time LLMs will have a noticeable effect on how fast non-ai scientific research is done, only time will tell.
AI in general, except some weirdness, has already surpassed humans in most thing for about a year or so, the fact is that AI is looped on TEXT DIFFUSION, humans have the resources and capabilities to research, use tools, loop through reasoning, etc. It's time for AI to expand to a multisensorial, memory, time regulated field of ideas that will turn a good LLM capable of doing even real world tasks without trying before. This is like Cambrian explosion for AI senses and external capabilities, and it will be a wild ride.
If you read X you will see how bad humans are at research
Hey Kyle, I don't really comment but I really have to say thank you for making these comparative deep analytical videos about the new AI's that come up especially things like Deep Research when things are actually coming close to the brim of being very relevant and useful for researchers and scholars alike.
I just want to say thank you for you to give your valuable time and expertise to give us a good real sense and understanding of where the current models stand and their significance. Truly, the whole community of STEM and even Non-STEM are all grateful for what you do and I hope you are well and keep continuing this work.
Thank you for your kind words, I really appreciate it!
I just found your channel. I would say the overall consensus in the comments section is that people enjoy your super deep and technical videos.
Make them as long as it takes.
You can create your own unique niche for hyper detailed and technical videos.
Great content!
Wow, Kyle, that was a deep dive! 🤯Seriously impressive how you put both AIs through the wringer with your own research. It's wild to see the difference in approaches. Gemini's approach reminds me of what I've been doing with my own AI agents: minimizing reliance on the LLM's pre-existing knowledge. For example, if I need a summary, I first have the agent research reputable sources on how to create the perfect summary. Then, it generates its own rules for summary creation, and only then applies those rules. That way, I'm not so dependent on the LLM's prior knowledge. Whether it knew how to summarize before or not, it's learned now, you know? It gives me way more freedom to choose smaller, faster models and helps ensure consistent quality, regardless of the LLM! 🚀Anyway, huge thanks for the thoroughness. Keep up the awesome work! 👍
wow, thanks for useful insight on agents with a smaller models onboard!
I saw this on Reddit some websites aren't actually paywalled, but they block web crawlers, making it seem like they are. I forgot the exact details, but I think it has to do with sites charging for access if you're using a web-scraping bot
A solution would be to use the chrome plugin to do the searching from the client device. That would also be better for geoblocked content (for open ai it would save a lot of bandwidth and compute time -> and better input for the deep research bot)
An interesting follow-up experiment would be to pass o3-mini-high your papers, assert that it must become the foremost authority on the papers, then evaluate and grade its prior research paper for accuracy, providing thorough and detailed reasoning, analysis and conclusions drawn from its findings.
Dr K really appreciate your expert contribution for evaluating these tools, there is so much info out there of non experts who hype up these things, and you can’t really get a true perspective. You provide that. Just keep it to 20 or 30 minutes like this one provides the most value.
That was a great video!
One wonders regarding the attempt to summarise the whole field is it’s just too much data for the model to deal with in one go. Maybe this could be done successfully if you asked o3 how to break it into manageable research topics, so summarising say 10 sub-fields.
Then giving those summaries to the Deep Research as attachments and asking it to produce a full summary. Just an idea.
Great idea!
@@KyleKabasares_PhD if it works, in theory it becomes possible to have an o3 Deep Research generated high quality encyclopaedia.
Thank you for your wonderful work. Looking forward to your video on Gemini 2.0 Flash Thinking.
Hi Kyle,
This video was much better. I understand that diving deep into your field can cause you to lose a large part of your audience. However, I believe most of your viewers are interested in seeing how deep AI can truly go. There is ongoing discussion and controversy regarding the capabilities and limitations of LLMs and their tool use. A structured approach, like the one in this video, is greatly appreciated. It would be fantastic to see an analysis on whether it followed the prompt, accessed the right sources, understood them, cited them accurately, identified or deduced controversial points, found any errors, and if the model itself made any errors.
Additionally, incorporating thoughts about evaluating the depth of research capabilities, assessing critical thinking skills in navigating complex data, and analyzing how effectively the model synthesizes information from various sources would further enhance the review process.
In my discussions with different thinking models, I discovered that after a while, the LLMs naturally seem to drift off and lose themselves in the woods. After a certain point, the quality of discussion breaks down to even fundamentals. This might be due to context window limitations or model cost. Anyway, I would be interested to see if anybody else has experienced this in other fields.
I see the same. I think this is the underlying reason no one can get agents to work effectively. Not sure whether it’s context window or lack of grounding in what it has done already. But you seem to have a limited depth to the conversation. Go to far and it ends up looping back.
Thank you! This is a great format for domain specific performance comparisons.
The problem with this experiment is that "Deep Research" was used as "Shallow Research." You could have achieved similar results by uploading the PDFs as context. It would also have been a good idea to revisit the list of requirements in the prompt and ensure the response actually addresses them.
Interesting test but imo you are limiting the capabilities of the research agent (deep research) in 2 ways: You are limiting how detailed it should answer 1500-2000 words. You are also telling it exactly where and what to search. This agent thrives on getting you information from the far corners of the internet. It finds the sources itself, gathering data, reasoning itself from source to source (using up to 100+ sources). I think thats the awesome part and it cant do that here.
I believe this task could have been done by any good reasoning model (deepseek, Gemini thinking, o1, o3) or even non-reasoning model with internet access (or those 2 PDFs attached).
Keep the tests coming, I find deep research really interesting and fascinating but its behind a big pay wall... for now.
Great video! This is the real purpose of AI-not just running random benchmarks or building a simple snake game.
Great review thanks, i wouldn't worry about being dull i think people are coming here to see how sophisticated the models are actually since benchmarks are not really reliable or standardized yet. Good stuff!
Did you give Gemini the links to your paper after it asked you about its research strategy? You can edit the strategy and tell it what and where to search and where not.
Great work!
1) can u prompt to find errors/mistakes in your work/articles/dissertation?))
2) can u prompt to do a particular part (sub-task) of your done work/article/dissertation (mb + python evaluator)?
To see, is it really useful by one-click right now)
Great video!! Very helpful to understand the capabilities of these newer research models. An excellent perspective offered by this experiment from an expert in this particular area of supermassive black hole astronomy.
Very useful. What I usually think about text GenAI is that they create phrases that in the context and style given, are most probable and suitable on that place. Maybe part of the human understanding is similar, we write with a motive and purpose which forms the message, but it cannot be just that. In a research one must have constant critical loop going on, which may cause a total turn in a direction, once new data and realization comes in. LLMs are transformers after all, turning one text to the other. But that said, these tools are impressive. Googles one probably requires a bit more experience and prompting how to get best of it. I wonder and got interested to try them with a more conceptual work in fundamental physics.
For me OpenAI Deep Research is like good and Brand new but Gemini 1.5 Deep Research It came in December, which is no surprise that it was bad for some people.
Overall: ChatGPT Deep Research> Gemini 1.5 Pro Deep Research
o3 mini high > Gemini 1.5 Pro
So waiting Gemini 2.0 Deep Research
And in Google Studio AI you could use the experimental version Gemini 2.0 Flash Thinking Exp
6:13 That might become something incredibly useful when Google implements their Titans architecture... 🤔
Your analysis is one of the view genuine results on the power of DeepResearch, thank you very much!
Thank you for making this video.
I haven’t tried OpenAi’s deep researcher, but found Google’s to be an absolute nothingburger.
It’s utterly broken, works one time out of 3-4, and when it works, it produces something on par with very basic o3-mini output. Just more wordy.
The quality of the answer is the quality of the question
very interesting. looks like it's really good. i hope they release it for plus tier soon
When you combine deep research and use those results as context for o1 or o1 pro magic happens.
That was a very good review. I found the length and amount of jargon to be fine - no problem. In fact, I would like to have learned more about your research, but of course that was not the topic. I was disappointed to learn that DeepResearch won't go behind paywalls. I thought for the price they charge they would include access to all the journals. Without that, it's not really that useful if it can only use open access sources. I was led to believe it would put a PhD to work for you, but what is the worth of a PhD without access to the journals?
Its actually unbelievable, It wrote a 12,000 word paper for me
I think Google is currently working on an exploration model that can reason like openai o3 but at present they are only using single blunt COT which is not enough for extracting nuances from query and information to provide better human level research info
Can we please have some timestamps?
Thanks for the informative update Kyle. Is deep research available for plus users on gpt as well
Open AI deep research is said to be powered by o3 fine tuned. I think that's why there is such a signifigcant different!
Kyle thank you
Thank you for watching!
Is Deep Thinking available for Gemini 2.0? I noticed you used 1.5, which is the previous "generation", so to speak.
Yes, it would be worth redoing this test with 2.0 Pro.
Could you ask AI to review this researched
article
Gemini deep research searches too much and confuses itself about what to include, I believe there is no reasoning behind what to search either and there is no way gemini 1.5 pro is in control of the search because I doubt it would search how to write papers, it's probably searching based on embeddings
Outstanding! Very helpful. Thank you.
I'm glad you found it helpful!
Fantastic video!
AI output is great if you know the field and can verify its output. This is good validation of the model.
Thanks for watching!
How does R1 Perform at this same task?
Take that test with Perplexity
its not gonna give answers nearly as deep and its not really for this kind of use case unless you wanna have a back and forth asking it different questions but its still a fantastic tool in its own right
I hope OpenAI watches these videos to learn how to improve
This is good but I still feel like it only tests a fraction of its power. I feel the best returns are when you have a set of partially or fully open-ended questions in the field and you specify a few base references to start with (they should be accessible to it) and then let it go off using its reasoning abilities. It's best to use a model like o3-mini or o1 to form a proper research plan. It's still not very suited for narrow research domains due to availability of material but really helpful when there are lot of resources on the internet.
Really interesting, thanks! :)
Very useful thanks!
Could use a bit of editing, removing the reading out loud. Awesome content!
Thanks for the tip!
i loved the video, thanks for sharing
Glad you liked it!
Deep research is interesting, but ultimately it feels like a information aggregator/collator rather than an "intelligent" agent..
Great video!
But Google's Is... ??
Really cool video!
It is not "o3 mini" deep research use the full o3
11:53 one time it has as reference (Kabasares et al. 2022) and under there (Kabasares et al. 2024) ist this wrong or right from the AI? It would wonder me if ChatGPT could now reference with the Harvard System
That was fantastic
Its also not $200/mo
I just can't stop laughing that you seem upset that Gemini's Deep Research isn't yet ready to replace you when it only costs $20/month. Shouldn't you be happy about that? Just think about all of the undergrads this can help find info or how much it can help someone who is checking to see if there is any information or research on a topic that is new to them. OMG, I cannot just give it a one sentence prompt and get it to write a paper for the Journal of Chemical Physics or something.
Why think when AI does it for you? Why work when AI does it for you? All you have left are your instincts. You return to your pre-human state.
"After leaving their bodies, they who have killed the Self go to the worlds of the Asuras, covered with blinding ignorance." - Upanishads.
Suras - gods, deities. World of suras - Heaven. Asuras - not suras. World of asuras - not Heaven, possibly hell.
Bard, Gemini, whatever you call it Google lost the AI race, accept it already!
$200 vs. $20, what do you expect?
Google is an advertiser using search as an interface, whereas OpenAI is a bona fide AI services provider, that's the difference, not the dollar amount.
@erkinalp If that's true, why doesn't OpenAI make the easy money from ads? Why do they charge $200?
@@gemini_537 you know not everyone has to do the business the same way even when pertaining to the same product or service, it's called business model and business niche
@@gemini_537because they believe they have enough value to provide with a sustainable subscription model to not have to stoop low enough to selling ads
@@gemini_537The same reason, Ferrari doesn't make the easy money and become a search company. Sit and think. The reason closedai is charging this much is it requires a lot of compute and their models are not efficient enough yet, as a for profit buisness, only 200$ can make then profitable alteast after some years. You are comparing a 2.5 trillion ads based search company vs a dedicated AI company worth 350 billion. Sit and think
This is the same as programming with AI at first it seems to generate code better than any human and everyone thinks is over, when you look closely you can see subtle mistakes and strange behaviors to the point you need human developers to fix the mess.
At the end of the day, you can make the AI say whatever whatever agenda you have.
Unless it has the feature like NotebookLM where you can see exactly in which line of the pdf file he got his knowledge, this will not be my deep research tool. i mean, ordinary web pages, really??? I'm a student and if we only used websites for reference, we'd be kicked out in less than 24 hours.
????
You mean "if we only used _Wikipedia_ for reference, we'd be kicked out in less than 24 hours."?
How about comparing openai-mini-high to Google Studio AI where you could use the experimental version Gemini 2.0 Flash Thinking Experimental 01-21 - gemini-2.0-flash-thinking-exp-01-21
🍓❤️
gemini loses
My friend, the best way to do what you're doing is to actually copy paste whole papers' text into the prompt. And then ask your questions. I prefer doing this in aistudio with gemini (don't use the Gemini app --- there are some potential quality differences, for scientific use I'm finding aistudio much better -- it's free as of Feb 2025). don't ask gemini to retrieve a document for you. Just do it yourself. Then ask your questions to their most powerful model (2.0 Pro model). Then you should get *very good / excellent* answers. You can do the same with o1 (or similar), btw. See what happens.
Gemini Pro 2.0: Okay, here's a report on the two provided research papers, structured in an academic style and addressing the key questions you've outlined. This version is designed to be easily copied and pasted into a Microsoft Word document without significant formatting problems.
Report: Dynamical Mass Measurements of Supermassive Black Holes Using ALMA and HST
Introduction
Supermassive black holes (SMBHs), with masses ranging from millions to billions of solar masses (M⊙), are believed to reside at the centers of most, if not all, massive galaxies. A key area of modern astrophysics is understanding the co-evolution of these SMBHs and their host galaxies. This co-evolution is suggested by observed scaling relations between the SMBH mass (MBH) and properties of the host galaxy's bulge, such as stellar velocity dispersion (σ⋆), luminosity (Lbul), and mass (Mbul) (e.g., Kormendy & Ho 2013). These relations imply a connection between the growth of the central black hole and the evolution of the galaxy itself.
To better understand these relationships and the underlying physics, it is crucial to obtain precise and accurate measurements of SMBH masses across a wide range of galaxy types and environments. Dynamical modeling, which involves analyzing the motion of stars or gas under the influence of the combined gravitational potential of the BH and the host galaxy, is a primary method for determining MBH.
The two papers under review here, Kabasares et al. (2022, hereafter Paper I) and Kabasares et al. (2024, hereafter Paper II), present dynamical measurements of SMBH masses in four early-type galaxies (ETGs): NGC 1380, NGC 6861, NGC 4786, and NGC 5193. They utilize observations from the Atacama Large Millimeter/submillimeter Array (ALMA) and the Hubble Space Telescope (HST) to model the kinematics of circumnuclear molecular gas disks.
Main Objective and Hypothesis
The primary objective of both research articles is to measure the masses of the central supermassive black holes in the target galaxies using gas-dynamical modeling of ALMA CO(2-1) observations, complemented by HST near-infrared imaging. The underlying hypothesis is that the observed molecular gas is in regular, dynamically cold rotation in a thin disk, and that its motion is primarily governed by the gravitational potential of the central BH and the stellar mass distribution of the host galaxy. By modeling this motion, the authors aim to constrain the BH mass.
Significance
These measurements are significant for several reasons:
Expanding the Sample: They increase the number of galaxies with dynamically measured BH masses, particularly in the ETG population. A larger sample allows for better statistical analysis of the scaling relations and reduces biases.
Testing Scaling Relations: The measurements provide independent checks on the MBH-σ⋆, MBH-Lbul, and MBH-Mbul relations, particularly at the high-mass end (for NGC 6861) and for galaxies with complex dust structures.
Probing Different Environments: The galaxies studied represent different environments (e.g., cluster vs. group), allowing for investigations into the potential influence of environment on BH growth.
Methodological Advancement: The papers refine and demonstrate the capabilities of ALMA for gas-dynamical BH mass measurements, particularly in challenging cases where the BH's sphere of influence (SOI) is marginally resolved or complicated by dust. They explicitly address systematic uncertainties, which are often underestimated.
Methods
Observational Data:
ALMA: Both papers utilize ALMA Band 6 observations of the CO(2-1) emission line.
Paper I: NGC 1380 (0.21" resolution), NGC 6861 (0.28" resolution).
Paper II: NGC 4786 (0.31" resolution), NGC 5193 (0.31" resolution).
The observations provide data cubes with spatial and spectral (velocity) information. The spectral resolution is crucial for resolving the gas kinematics.
The observations were chosen to target galaxies with known circumnuclear dust disks, which often harbor molecular gas.
HST: Archival HST/WFC3 near-infrared (NIR) imaging in the F110W (J-band) and F160W (H-band) filters was used.
The NIR imaging is less affected by dust extinction than optical imaging, providing a better view of the stellar distribution.
The images were processed (drizzled, aligned) to create high-quality mosaics.
J-H color maps were constructed to identify regions of significant dust attenuation.
Host Galaxy Modeling:
A crucial step is to model the stellar mass distribution of the host galaxy, as this contributes significantly to the gravitational potential. The authors used the following approach:
Multi-Gaussian Expansion (MGE): The HST H-band images were modeled using the MGE method (Emsellem et al. 1994; Cappellari 2002). This involves fitting a series of two-dimensional Gaussian functions to the observed surface brightness distribution. The MGE provides an analytical representation of the galaxy's light profile.
Dust Correction: Because dust significantly affects the observed surface brightness, especially in the central regions, the authors implemented several strategies:
Dust Masking: Regions with high J-H color excess (indicating dust) were masked out before fitting the MGE.
Dust Correction: A more sophisticated approach involved using a Nuker model (Faber et al. 1997) to fit the central surface brightness profile and then correcting for dust extinction using an embedded-screen model (Viaene et al. 2017) and the observed J-H color. This allowed for an estimate of the intrinsic (unobscured) stellar light profile. Different levels of extinction were tested to assess the systematic uncertainty.
Deprojection: The 2D MGE models were deprojected to obtain a 3D representation of the stellar luminosity density, assuming the galaxies are oblate and axisymmetric. This deprojection is necessary to calculate the stellar contribution to the circular velocity.
Gas-Dynamical Modeling:
The core of the analysis is the gas-dynamical modeling, which involves the following steps:
Thin Disk Model: The molecular gas is assumed to reside in a thin, rotating disk. This is a reasonable assumption given the observed morphology and kinematics of the CO emission.
Circular Velocity: The circular velocity (vc) of the gas at a given radius (r) is calculated based on the enclosed mass:
vc(r) = √(GMBH/r + ΥH * v2
⋆,MGE(r) + v2
gas(r))
MBH: Black hole mass (a free parameter).
ΥH: Stellar mass-to-light ratio in the H-band (a free parameter).
v⋆,MGE: Circular velocity due to the stellar mass distribution (derived from the deprojected MGE).
vgas: Circular velocity due to the gas mass distribution (estimated from the CO flux).
Model Cube Generation: A model data cube is generated, mimicking the ALMA observations. For each point in the disk:
The LOS velocity is calculated based on the circular velocity, disk inclination (i), and position angle (Γ) (free parameters).
A Gaussian line profile is assumed, with the centroid determined by the LOS velocity and the width determined by a turbulent velocity dispersion (σ0) (a free parameter).
The line profile is weighted by a CO flux map, derived from the ALMA data (deconvolved using the Richardson-Lucy algorithm).
Beam Convolution: The model cube is spatially oversampled (typically by a factor of 3) to better model steep velocity gradients. However, the computationally expensive beam convolution (to match the ALMA beam) is performed after downsampling the model to the original ALMA pixel scale.
χ2 Minimization: The model cube is compared to the ALMA data cube, and a χ2 statistic is calculated. A Levenberg-Marquardt algorithm is used to minimize χ2 by adjusting the free parameters (MBH, ΥH, i, Γ, xc, yc, F0, σ0, vsys). The fitting is performed over a defined spatial region (typically an ellipse encompassing the inner part of the disk) and a range of frequency channels.
Noise Model: The authors employed a noise model that accounted for the spatial variation in the background noise introduced by the primary beam of the ALMA telescope.
Results
The measured BH masses are summarized below:
Paper I:
NGC 1380: MBH = 1.47 × 108 M⊙ (with a ~40% uncertainty, dominated by systematics). The range due to differing dust corrections is (1.02 − 2.04) × 108 M⊙.
NGC 6861: MBH = (1 − 3) × 109 M⊙ (a range reflecting the large uncertainty due to the central hole in the gas distribution).
Paper II:
NGC 4786: (MBH/108 M⊙) = 5.0 ± 0.2 [1σ statistical] +1.4
−1.3 [systematic] ± 0.75 [distance].
NGC 5193: (MBH/108 M⊙) = 1.4 ± 0.03 [1σ statistical]+1.5
−0.1 [systematic] ± 0.1 [distance].
Major Sources of Uncertainty
The authors carefully considered and quantified various sources of uncertainty:
Dust Extinction: This is the dominant source of systematic uncertainty in both papers. The correction for dust attenuation in the host galaxy model significantly impacts the inferred stellar mass profile and, consequently, the BH mass. The authors explored a range of extinction corrections, leading to a substantial spread in the derived MBH values.
Host Galaxy Model: Related to the dust issue, the choice of the host galaxy model (e.g., unmasked, dust-masked, dust-corrected MGE) introduces significant uncertainty, particularly when the BH SOI is not well resolved. Different models lead to different inner stellar mass profiles.
Fit Region: The choice of the spatial region used for fitting the models can influence the results. Fitting the entire disk can be sensitive to uncertainties in the outer stellar mass profile, while fitting a smaller region can be more affected by beam smearing.
Gas Mass: The contribution of the gas mass to the total potential is relatively small, but its inclusion or exclusion can have a minor effect on MBH. The adopted αCO value to convert CO luminosity to gas mass has an inherent uncertainty.
Radial Motion: Deviations from pure circular motion (e.g., radial inflow) are possible, though the authors found that their models strongly favored pure rotation.
Turbulent Velocity Dispersion: While the disks are generally dynamically cold, the authors tested models with a spatially varying turbulent velocity dispersion, finding little impact on MBH.
Numerical Parameters: Parameters like the pixel oversampling factor and block-averaging factor were tested and found to have relatively minor effects.
Distance Uncertainty: The assumed distance to the galaxy directly scales the derived MBH. The distance uncertainties contribute to the overall error budget.
Statistical Uncertainty: Formal statistical uncertainties (derived from Monte Carlo simulations) are generally much smaller than the systematic uncertainties.
Discussion
The papers demonstrate the power of ALMA for measuring SMBH masses through gas-dynamical modeling. However, they also highlight significant challenges:
Resolution: Ideally, the BH SOI should be well resolved to minimize the degeneracy between the BH and stellar mass. When the SOI is marginally resolved or unresolved (as in NGC 5193 and to a lesser extent NGC 1380 and NGC 4786), the uncertainty in the host galaxy model becomes critical.
Dust: Dust extinction is a major obstacle, particularly in ETGs, which often have prominent dust disks. Accurate dust correction is crucial for obtaining a reliable stellar mass profile. The simple embedded-screen model used by the authors is a useful approximation, but more sophisticated radiative transfer modeling would be ideal.
Central Holes: The presence of a central hole in the gas distribution (as in NGC 6861) severely limits the ability to constrain MBH, as there are no kinematic tracers within the SOI.
Comparison to Scaling Relations
The measured BH masses were generally consistent with predictions from the MBH-Lbul and MBH-Mbul relations, but tended to fall below the predictions of the MBH-σ⋆ relation, particularly for NGC 6861. The tension between MBH predictions from different scaling relations has been recognized as a critical issue in the field.
Conclusion
Kabasares et al. (2022, 2024) present valuable new measurements of SMBH masses in four ETGs, using ALMA and HST data. The work highlights both the capabilities and the limitations of gas-dynamical modeling with ALMA. While ALMA provides excellent spatial and spectral resolution, the presence of dust and the limited resolution relative to the BH SOI in some cases introduce significant systematic uncertainties. The authors' careful consideration of these uncertainties, particularly those related to dust extinction, is a strength of the papers.
Areas for Improvement
The authors acknowledge several areas where future work could improve the measurements:
Higher Resolution ALMA Observations: Obtaining higher-resolution ALMA data would better resolve the BH SOI, reducing the degeneracy between BH and stellar mass.
Deeper ALMA Observations: Deeper observations could potentially detect fainter CO emission closer to the center, providing more kinematic constraints within the SOI.
3D Radiative Transfer Modeling: Implementing 3D radiative transfer models (e.g., De Geyter et al. 2013; Camps & Baes 2015) would provide a more accurate treatment of dust extinction and scattering, leading to more reliable host galaxy models. This is computationally expensive but would significantly reduce the dominant source of systematic uncertainty.
Multi-Wavelength Data: Combining ALMA data with observations at other wavelengths (e.g., near-infrared integral field spectroscopy) could provide additional constraints on the stellar population and dust properties.
Modeling Non-Circular Motions: While the authors found little evidence for significant non-circular motions, more sophisticated models that allow for warps or other deviations from a perfectly thin, rotating disk could be explored.
Stellar Dynamical Modeling: Complementary modeling efforts using stellar dynamics, could provide an alternative constraint on the enclosed mass.
Υ Gradients: Future efforts can account for potential gradients in stellar mass-to-light ratios that are not considered in this work.
In summary, these papers represent important contributions to the field of SMBH mass measurements and highlight the ongoing challenges and opportunities in understanding the co-evolution of black holes and galaxies. The explicit treatment of systematic uncertainties is a valuable aspect of this work.