LessWrong 2.0 Reader

I'm claiming that even if you go all the way to BoN, it still doesn't necessarily leak less info to the morel

Something between training the whole model with RL and BoN is training just the last few layers of the model (for current architectures) with RL and then doing BoN on top as needed to increase performance. This means most of the model won't know the information (except insofar as the info shows up in outputs) and allows you to get some of the runtime cost reductions of using RL rather than BoN.

for a sufficiently competent policy, the fact that BoN doesn't update the policy doesn't mean it leaks any fewer bits of info to the policy than normal RL

I'm just talking about research intended to be safety/safety-adjacent. As in, of this research, what has the quality weighted differential safety progress been.

Probably the word "differential" was just a mistake.

I've been thinking that EA should try to elect a president, someone who is empowered but also accountable to the general people in the movement, a schelling person to be the face of EA.

Counterargument, I think there's enough different streams of EA that this would not be especially helpful.

There exists a president of GiveWell. There exists a president of 80k Hours. There exists a president of Open Philanthropy. Those three organizations seem pretty close to each other, and there's a lot of others further afield. I think there would be a lot of debating, some of it acrimonious, about who counted as 'in the movement' enough to vote on a president of EA, and it would be easy to wind up with a president that nobody with a big mailing list or a pile of money actually had to listen to. 

I think Steve Hsu has written some about the evidence for additivity on his blog (Information Processing). He also talks about it a bit in section 3.1 of this paper.

This is a crux. I acknowledge I probably share more values with a random EA than a random university student, but I don't think that's actually saying that much, and I believe there's a lot of massively impactful difference in culture and values.

My best guess is something like a third of rationalists are also EAs, at least going by identification [? · GW]. (I'm being lazy for the moment and not cross checking "Identifies as Rationalist" against "Identifies as EA" but I can if you want me to and I'm like 85% sure the less-lazy check will bear that out.) My educated but irresponsible guess is something like 10% of EAs are rationalists [? · GW]. Last time I did a straw poll at an ACX meetup, more than half the people attending were also EAs. Whatever the differences are, it's not stopping a substantial overlap on membership, and I don't think that's just at the level of random members but includes a lot of the notable members.

I'd be pretty open to a definition of 'rationalist' that was about more than self-identification, but to my knowledge we don't have a workable definition better than that. It's plausible to me that the differences matter as you lean on them a lot, but I think it's more likely the two groups are aligned for most purposes. 

As other people have said, this is a known argument; specifically, it's in The Generalized Anti-Zombie Principle [LW · GW] in the Physicalism 201 [? · GW] series. From the very early days of LessWrong

Albert: "Suppose I replaced all the neurons in your head with tiny robotic artificial neurons that had the same connections, the same local input-output behavior, and analogous internal state and learning rules."

I think this proof relies on three assumptions. The first (which you address in the post) is that consciousness must happen within physics. (The opposing view would be substance dualism where consciousness causally acts on physics from the outside.) The second (which you also address in the post) is that consciousness and reports about consciousness aren't aligned by chance. (The opposing view would be epiphenomenalism, which is also what Eliezer trashes extensively in this sequence.)

The third assumption is one you don't talk about, which is that switching the substrate without affecting behavior is possible. This assumption does not hold for physical processes in general; if you change the substrate of a plank of wood that's thrown into a fire, you will get a different process. So the assumption is that computation in the brain is substrate-independent, or to be more precise, that there exists a level of abstraction in which you can describe the brain with the property that the elements in this abstraction can be implemented by different substrates. This is a mouthful, but essentially the level of abstraction would be the connectome -- so the idea is that you can describe the brain by treating each neuron as a little black box about which you just know its input/output behavior, and then describe the interactions between those little black boxes. Then, assuming you can implement the input/output behavior of your black boxes with a different substrate (i.e., an artificial neuron), you can change the substrate of the brain while leaving its behavior intact (because both the old and the new brain "implement" the abstract description, which by assumption captures the brain's behavior).

So essentially you need the neuron doctirne to be true. (Or at least the neuron doctrine is sufficient for the argument to work.)

If you read the essay, Eliezer does mostly acknowledge this assumption. E.g., he talks about neuron's local behavior, implying that the function of a neuron in the brain is entirely about its local behavior (if not, this makes the abstract description more difficult at least, may or may not still be possible).

He also mentions the quantum gravity bit with Penrose, which is one example of how the assumption would be false, although probably a pretty stupid one. Something more concerning may be ephaptic coupling, which are non-local effects of neurons. Are those copied by artificial neurons as well? If you want to improve upon the argument, you could discuss the validity of those assumptions, i.e., why/how we are certain that the brain can be fully described as a graph of modular units.

(Also, note that the argument as you phrase it only proves that you can have a brain-shaped thing with different neurons that's still conscious, which is slightly different from the claim that a simulation of a human on a computer would be conscious. So if the shape of the brain plays a role for computation (as e.g. this paper claims), then your argument still goes through but the step to simulations becomes problematic.)

links 10/25/24: https://roamresearch.com/#/app/srcpublic/page/10-25-2024

• https://theoryandpractice.org/2024/10/Yes,%20we%20did%20discover%20the%20Higgs!/  CERN's statistical methods are good actually. compare this to any other stats-heavy area of natural or social science and they come out impressively rigorous. blinded data analyses? whoa.
• https://en.m.wikipedia.org/wiki/Siege_engine  a siege engine is any machine you use against the city you're besieging -- from towers to catapults to flamethrowers to artillery.
• there's a new field of "pan-cancer" where you make (mostly molbio) comparisons across cancers, including vulnerability screens where you use CRISPR or RNAi to knock down each gene and see which ones kill the cancer cells when absent.
  • https://aacrjournals.org/clincancerres/article/24/9/2182/81290/Pan-Cancer-Molecular-Classes-Transcending-Tumor representative paper
  • https://www.nature.com/articles/s41467-019-13528-0 you can also do it with the proteome
  • https://www.nature.com/articles/s41422-020-0355-0 you can single-cell it
  • https://www.science.org/doi/10.1126/science.abe6474 you can profile the TILs cell by cell
  • https://link.springer.com/content/pdf/10.1186/s13059-023-03020-w.pdf CRISPR and RNAi both have their strengths and weaknesses but if you look at the overlap there are still a bunch of "pan-essential" genes that all cancers need to survive. (do healthy cells also need those, or are they good therapeutic targets? we just don't know.)
  • https://www.researchgate.net/profile/Yize-Li/publication/354641293_Moving_pan-cancer_studies_from_basic_research_toward_the_clinic/links/615f5d570bf51d4817512465/Moving-pan-cancer-studies-from-basic-research-toward-the-clinic.pdf?_sg%5B0%5D=started_experiment_milestone&_sg%5B1%5D=started_experiment_milestone&origin=journalDetail by "towards the clinic" we mean "very gingerly", apparently
  • https://www.cell.com/cancer-cell/fulltext/S1535-6108(20)30656-5 when you target pan-essential genes you are in the chemo zone, where by default the therapeutic index is low (kills cancer AND healthy cells) and you need to put more work in to handling toxicity.
• tumors -- most tumors -- get coated with IgG, much more than other tissues. is this a systemic defense against cancer or a tumor-secreted IgG? hard to say.
  • https://www.cell.com/cell/pdf/S0092-8674(22)00192-1.pdf mostly ovarian carcinoma but they also compare to a bunch of other tumor types
  • https://www.mdpi.com/1422-0067/22/21/11597 could be tumor-derived
  • https://www.nature.com/articles/srep05088.pdf this is Sanford Simon -- endogenous IgG concentrates around mouse tumors of many types
• James Watson's vision for cancer research -- this is what he originally became a "controversial figure" for, before the race thing. https://royalsocietypublishing.org/doi/full/10.1098/rsob.120144
  • basically, this is a few things:
    • a call to work harder dammit and treat it like a true war on cancer, not a sedate and bureaucratic academic field
    • a call for more pan-cancer RNAi vulnerability screens
    • a call to focus on transcription factors as targets, particularly things like Myc and BRD4 that are particularly involved in the transition to metastasis -- we don't yet have any good drug therapies that work well on metastatic cancers
      • transcription factors are obviously causally upstream of what makes cancer cancer -- its invasiveness, its metastatic potential, its evasion of immune surveillance, etc
      • they are hard to drug though, because they're in the nucleus, not on the cell surface. but we can start to do hard things now!
      • cell surface growth factors (think EGFR) are the easiest to target but the associated drugs have unimpressive clinical effects in most patients because targeting growth factors only slows growth, it doesn't kill cancer cells. usually just slightly delays the inevitable.
    • a statement of his redox hobbyhorse -- ROS is good, ROS is how the body fights cancer, etc.
      • not sure how to operationalize this as a strategy. it might, as it turns out, be redundant with immunotherapy.
    • a couple specific targets/mechanisms he thinks deserve more attention -- apparently the circadian regulator PER2 is a tumor suppressor. i'm always down for more attention to circadian stuff.
• the Halifax Project researched the hypothesis that low-dose combinations of environmental carcinogens might synergistically increase cancer risk:
  • https://www.ewg.org/research/rethinking-carcinogens
  • https://www.degruyter.com/document/doi/10.1515/reveh-2020-0033/html
• https://en.wikipedia.org/wiki/CpG_oligodeoxynucleotide this is the inflammatory molecule on bacteria that's the reason bacterial infections sometimes cause complete regressions of very difficult tumors (like sarcoma -- we have no drugs for sarcoma! it's either surgery or death!). fortunately the immunooncology people are On It and researching this as an immunostimulant.
  • if you've heard of "Coley Toxins", they're kind of an alt-med thing with a tantalizing grain of truth -- but we don't need to inject bacteria into tumors any more, we know how they work, we can replicate the effect with well-defined compounds now.
• https://www.cell.com/cell-chemical-biology/fulltext/S2451-9456(23)00221-0?rss=yes#mmc1 this is AOH1996, the mindblowingly selective new pan-cancer drug candidate.
  • basically this is using the same principle as old-fashioned chemo -- hit it in the DNA replication -- but with a new target, and with modern structural-biology-based rational drug design to hit the cancer version of the target rather than the healthy-cell type.
    • would I have guessed there was room for optimism here a priori? no way.
    • but apparently we have not explored this space sufficiently. now try it with AlphaFold.
    • https://www.science.org/content/blog-post/new-mode-cancer-treatment and Derek Lowe is impressed.
    •  
• https://www.nature.com/articles/s41419-021-04468-z inducible caspase 9 allows conditional apoptosis. it's incredibly powerful. unfortunately it doesn't always work and this raises drug resistance concerns.
  • I haven't yet seen many examples of "put the iCasp9 in the cell if-and-only-if the cell has some molecular marker" but that's the obvious place to go.
  • you can kinda reduce the drug resistance thing by putting a promotor to increase iCasp9 expression. buddy if this is where we are in 2022 i'm going to predict there is a LOT of potential value in continuing to work out the kinks in this system. get in on the ground floor!
• https://en.wikipedia.org/wiki/Proteolysis_targeting_chimera a PROTAC is "this protein? kill it." uses the ubiquitin system.
• sadly, the Warburg Effect is not as cool as I once thought.
  • glucose deprivation is just not that deadly across tumor lines: https://www.sciencedirect.com/science/article/abs/pii/S0899900720300319
  • https://pmc.ncbi.nlm.nih.gov/articles/PMC3237863/ i mean, 2DG + metformin might do a thing?
  • https://pmc.ncbi.nlm.nih.gov/articles/PMC5095922/#BST-2016-0094C20 yeah...it's not what you think.
  • https://www.sciencedirect.com/science/article/abs/pii/S0006291X0302504X if you actually measure what % of ATP comes from glycolysis, cancers cover a wide range, and the distribution overlaps substantially with the distribution of healthy cells. glycolysis dominance is not a distinguishing characteristic of all or even most cancers.
  • https://journals.sagepub.com/doi/pdf/10.2310/7290.2015.00021 heavy glucose uptake is enough of a thing for PET imaging to be used clinically though
  •  
• are cancer cells selectively vulnerable to mechanical stress? kinda, but also it sometimes stimulates them to go metastatic so beware.
  • https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/full/10.1002/elsc.201900154 vibrate em and they apoptose
  • https://www.sciencedirect.com/science/article/abs/pii/S0168365915302819 just mcfuckin spin some tiny magnets around in your brain tumor. apparently it works in rodents but aaaaaah
  • https://www.sciencedirect.com/science/article/pii/S0167488913003224 laminar but not oscillatory shear stress kills em? no non-cancer comparison
  • https://www.sciencedirect.com/science/article/pii/S2949907024000585 vibrations to kill prostate cancer? no non-cancer comparison
  • https://pubs.rsc.org/en/content/articlelanding/2015/nr/c5nr03518j/unauth magnetic particle vibration against renal cancer? no non-cancer comparison
  • https://www.nature.com/articles/s41557-023-01383-y.epdf?sharing_token=jICYt2mKBMQ0GsetiWodv9RgN0jAjWel9jnR3ZoTv0PPtLuduvirY9e9lvJJx5Q_iJTfP9UCvLlXVOkNBly5J-gi3DlHLxMYWqsmEJBOrH0s7RbtQm1UREc3FbrfF2vDNLzTfS250KEAwBdVsczhxamax0pSp4TP23jM_ehG703560use7dJ6hnsaVLpnXsWU1n14UplHLGvaXHsJ444z96C3IEcjmnjMZvijAgkKsQ%3D&tracking_referrer=www.genengnews.com "vibronic molecular jackhammers"? still no non-cancer comparison, but at least they tried some mice
  • https://www.nature.com/articles/s41413-020-00111-3 vibration to reduce metastatic potential of breast cancer cells. no non-cancer comparison.
  • https://www.science.org/doi/10.1126/science.adp7206 gently ultrasound the tumor to sensitize it to chemo or induce an immune anti-tumor response.
  • https://pmc.ncbi.nlm.nih.gov/articles/PMC10068349/ Piezo1 might be involved in an apoptotic response to mechanical stress?
  • https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/full/10.1002/elsc.201900154 vibrate some cancer cells and they go apoptotic but not necrotic. no non-cancer comparison.
  • https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4703126 more ultrasound, including in vivo, Piezo1 mediated.
  • https://pmc.ncbi.nlm.nih.gov/articles/PMC8274378/ mechanical stress is also a natural feature of cancer -- tumors get more rigid and experience pressure. in fact this stress can be a trigger for increased proliferation or metastasis, so watch out!
  • https://pmc.ncbi.nlm.nih.gov/articles/PMC5992512/ oops shear stress can promote metastasis
  • https://www.cell.com/biophysj/fulltext/S0006-3495(22)00367-8 substrate stiffness promotes invasion and metastasis
  • https://www.frontiersin.org/journals/pharmacology/articles/10.3389/fphar.2022.955595/full "mechanoptosis" (mechanical pressure causing cancer apoptosis)
  • https://elifesciences.org/for-the-press/12916d1e/migrating-through-small-spaces-makes-cancer-cells-more-aggressive squish cancer cells through tight spaces (eg on a microfluidic chip) and you get more invasive/metastatic potential
  •  
• are cancer cells selectively vulnerable to electrical stress? also kinda yeah
  • https://www.mdpi.com/2072-6694/13/9/2283 "tumor treating fields", just an oscillating electric field, are actually an approved therapy in glioblastoma that extends life a few months. (not saying much though...glioblastoma is so deadly that it's easy mode from an FDA standpoint)
  • of course you can just kill *cells* with pulsed electric fields, cancer or not: https://faseb.onlinelibrary.wiley.com/doi/abs/10.1096/fj.02-0859fje
  • https://jamanetwork.com/journals/jama/article-abstract/2475446 more electrical fields for glioblastoma
  • https://aacrjournals.org/cancerres/article/64/9/3288/517864/Disruption-of-cancer-Cell-Replication-by ah this actually IS a differential effect in tumor vs. non cancer cell lines. plus in vivo, in mice.
• i don't even know man. somebody who knows physics explain this. little nanoelectrodes with some chemical functionalization kill cancer cells? "quantum biological tunneling?" https://www.nature.com/articles/s41565-023-01496-y
• https://nyulangone.org/news/coping-mechanism-suggests-new-way-make-cancer-cells-more-vulnerable-chemotherapies "stress granules" as a form of chemo resistance, driven by KRAS?
• https://www.nature.com/articles/s41420-022-01202-2 cancer can be selectively vulnerable to proteotoxic stress. they're worse at expressing heat shock proteins.
  • https://www.researchgate.net/profile/Pietro-Taverna-2/publication/221748736_The_Novel_Oral_Hsp90_Inhibitor_NVP-HSP990_Exhibits_Potent_and_Broad-spectrum_Antitumor_Activities_In_Vitro_and_In_Vivo/links/56d0a18708ae059e375d4920/The-Novel-Oral-Hsp90-Inhibitor-NVP-HSP990-Exhibits-Potent-and-Broad-spectrum-Antitumor-Activities-In-Vitro-and-In-Vivo.pdf heat shock protein inhibitor reduces tumor growth in many cell lines
• cancer cells have depolarized membranes -- you can literally distinguish them from healthy cells by voltage alone.
  • this is a Michael Levin thing. https://pmc.ncbi.nlm.nih.gov/articles/PMC3528107/ you can give a frog a tumor -- or make the tumor go away -- through manipulating voltage alone! it does not matter what ion channel you use, it's about the voltage.
    • more Michael Levin https://pmc.ncbi.nlm.nih.gov/articles/PMC4267524/#R250
  • apparently Wnt signaling is involved. https://physoc.onlinelibrary.wiley.com/doi/abs/10.1113/JP278661 in general you get alterations in membrane voltage potential by changing the behavior of ion channels
  • https://jamanetwork.com/journals/jamasurgery/article-abstract/591620#google_vignette depolarization occurs early in the development of colon cancer in mice exposed to a carcinogen.
  • https://www.medigraphic.com/pdfs/hepato/ah-2017/ah172s.pdf cancer stem cells are depolarized relative to normal stem cells
  • https://www.nature.com/articles/s41598-021-92951-0.pdf here's math modeling if you care.
  • https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2013.00185/full a lot of ion channels are involved
  • https://karger.com/tbi/article-abstract/15/3/147/299607/Electrical-Potential-Measurements-in-Human-Breast breast cancers vs non-cancer tumors show up differently on an external volt meter!!!!
    • literally, you can try this at home! stick a voltmeter across your boob!
  • https://aacrjournals.org/cancerres/article/40/6/1830/484668/Cellular-Potentials-of-Normal-and-cancerous cancerous cells have lower membrane potential than their healthy counterparts
  • https://nyaspubs.onlinelibrary.wiley.com/doi/abs/10.1111/j.1749-6632.1974.tb26808.x even in non-cancer cells membrane potential correlates negatively with proliferation
  • https://pmc.ncbi.nlm.nih.gov/articles/PMC9652252/ cancer membrane potentials also fluctuate more than healthy cell membrane potentials
  • https://aacrjournals.org/amjcancer/article/32/2/240/679553/Bio-Electric-Properties-of-Cancer-Resistant-and going back to 1938, if you put a volt meter across a mouse's body you can tell the ones with tumors from the ones without. it is literally that simple and has been known that long.
• something in (some of?) the neutrophils in (some) humans and a cancer-resistant strain of mice can kill cancer, including when transferred. a Zheng Cui research program.
  • my take is, he's not an immunologist and modern methods could elucidate the specific clonal population a LOT better than this, but I like the thought process.
  • https://www.cell.com/heliyon/fulltext/S2405-8440(17)31693-6 they did some infusions from young blood donors into 3 patients with advanced metastatic cancer got a bunch of tumor necrosis and a cytokine release syndrome. all died within 3 months though.
  • https://link.springer.com/article/10.1186/1475-2867-11-26 healthy controls' leukocytes are better at killing cancer in vitro than cancer patients'. this is as expected.
  • https://link.springer.com/article/10.1186/1471-2407-10-179 SR/CR cancer resistant mice seem to need the leukocytes to physically home to the cancer cells. again, not news; neutrophils infiltrate tumors.
    • Cui has been beating this drum since before immunotherapy was cool, so let's not blame him too much, but we do very much know this bit independently
    • eg https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2019.01710/full -- sometimes neutrophils promote cancer actually!
    • eg https://www.cell.com/cell-reports/pdfExtended/S2211-1247(22)00984-6 we can determine the "good guy" neutrophil subpopulation that infiltrates tumors and promotes an anti-tumor immune response: it's HLA-DR+CD80+CD86+ICAM1+PD-L1-. in metastasis these guys become PD-L1+ and immunosuppressive.
    • so like...the secret to replicating Zheng Cui's miracle mice...might be nivolumab?? don't get me wrong it's a good drug but this is anticlimactic.
      • a great example of the mundanity of success
  • https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0059995&type=printable the Danes do not replicate quite as much cancer resistance from SR/CR mice as Cui's lab
  • https://www.pnas.org/doi/pdf/10.1073/pnas.0602382103 whatever the SR/CR mice are doing, you can transfer it to other mice and get cancer resistance. Lloyd J. Old is a coauthor!!!
  • https://pmc.ncbi.nlm.nih.gov/articles/PMC4544930/ independent description of anti-tumor neutrophils extracted from mice
  • https://www.sciencedirect.com/science/article/abs/pii/S0171298510000033 more evidence of anti-tumor granulocytes(which include neutrophils)
  • https://www.tandfonline.com/doi/abs/10.4161/cbt.7.9.6417 cancer patients' granulocytes are less active
• more Zheng Cui: cancer cells are negatively charged, such that positively charged nanoparticles can detect them VERY specifically. this is legit IMO.
  • https://link.springer.com/article/10.1007/s41048-018-0080-0 works on 22 different cancer cell lines. absolutely no affinity for healthy cells, quite a bit for all cancer cells.
  • https://www.thno.org/v06p1887.htm
    • can't do it in vivo though
  • https://link.springer.com/article/10.1186/s12951-019-0491-1?fromPaywallRec=false detects four CTCs per 1 mL blood!!!!
  • https://www.science.org/doi/10.1126/science.abm5551#supplementary-materials sadly these guys think positive nanoparticles are too toxic to use as treatments -- the entire paper is about negative nanoparticles, which do sometimes add to the tumor uptake of chemotherapies
• https://www.sciencedirect.com/science/article/abs/pii/S1879625715000413 oncolytic virus BHV1 kills cancer cells in a variety of tumor types?
• https://www.nature.com/articles/s41598-023-47478-x.pdf broad-spectrum metastasis suppressing compounds targeting a lncRNA. scary-big chemical structures though.
• https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1349-7006.2010.01834.x broad spectrum effectiveness of a survivin inhibitor + tumor regressions in vivo.
  • survivin is pan-essential, i got a good feeling about this
• https://aacrjournals.org/cancerimmunolres/article/2/6/510/467367/VISTA-Is-a-Novel-Broad-Spectrum-Negative VISTA is another negative checkpoint regulator like PD1 and CTLA4 (which are both major successful drug targets)
• https://www.science.org/doi/abs/10.1126/scitranslmed.3007646 alkylphosphocholine is a type of lipid especially present in cancer cells, across cancer types, via lipid rafts. a synthetic analog has preferential uptake in basically all rodent & human tumors. usable for imaging and radiotherapy.
• https://www.nature.com/articles/s41568-023-00554-w Sanford Simon's personal journey against fibrolamellar hepatocellular carcinoma
  • they found a fusion transcript and a corresponding fusion protein -- the root cause
  • they did the reasonable thing: screen a compound library against tumor samples.
    • one hit is napabucasin, usually known as a STAT3 inhibitor (but that's not the mechanism here) but somebody owns it
    • another was irinotecan. and navitoclax...but navitoclax has platelet toxicity
      • irinotecan + a BcrX PROTAC is being investigated though
    • or you can just. shRNA the fusion transcript. that's a thing you can do now.
      • apparently Elana wanted to do that in 2013 but her dad said "pshaw RNA breaks down in the body." now Spinraza is a thing (antisense oligonucleotide.) not to mention the mRNA world. truly these are the days of miracle and wonder.
• https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2018.00126/full redox balance is tricky, since cancers are both more prone to ROS and prone to develop coping mechanisms to regain redox balance.
• https://www.nature.com/articles/cdd2017180 there's nothing like TP53. mutated in 50-60% of cancers. the tumor suppressor gene par excellence.
• https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3-Technical-Report.pdf the Hermes 3 model is fine-tuned to be more responsive to prompts, such that prompt engineering suffices for "in-character" writing style (which IME does not work on most Instruct models)
•  

This is pretty well-trodden ground.  The scalable hiring/matching paths are effective enough and cheap for the bulk of the bell curves (of positions and of candidates).  If you're pretty normal, or seeking a pretty normal candidate, that's fine.  

If you're exceptional, or if you're a startup looking for exceptional employees, then it's not going to work very well.  Tyler Cowen's Talent (Talent: How to Identify Energizers, Creatives, and Winners Around the World: Cowen, Tyler, Gross, Daniel: 9781250275813: Amazon.com: Books) makes this point very clearly.  

It's kind of sad that it's so difficult to fire people nowadays, because it makes it very risky to hire someone that you're not pretty confident won't be terrible.  It changes the dynamic from "take risks, hire high-potential" to "avoid risks, hire the safe".  And this really sucks in a lot of ways.