Flagship Afterthoughts: Could the Human Brain Project (HBP) Have Done Better?

Building a new scientific community

A crucial hurdle, from the start, was that there was no solid census of scientists who would not only be interested enough, but who would also give their full adherence to a collective research roadmap. Science is often done by bright minds obsessed with private/individual narrowly-focused agendas. Apart from some notable mid-size FP7 and FET-Open initiatives (Daisy, BrainScales, Robocom), which gave rise to distinct and somewhat opposing lobbying (Dario et al., 2011; Markram et al., 2011; see also Martin, 2022), very few structuration efforts existed in Europe at a macroscale level before HBP. An attempt to consolidate scientific networks was made as early as 2004 by the National Bernstein Computational Neuroscience initiative. Its aim was to establish central nodes of excellence to strengthen German regional capacities in the field of computational neuroscience, interconnect them and develop applications in the neuro-IT and brain-machine interface domains. This structuring move was the result of a unique, large-scale funding initiative (in sum over 200 million euros) of the German Federal Ministry of Education and Research (BMBF), led with persistence by a remarkable officer, Christiane Buchholz (Schwarzwaelder and Cardoso de Oliveira, 2010; Biopro, 2011), and driven by ardent and dedicated scientists (Andreas Herz and Ad Aertsen, among others). This impressive success story, unique in the European scientific landscape in life sciences, should have inspired HBP. However, unfortunately for HBP, the internal conflict between theoreticians and Henry Markram, during the writing of the roadmap in 2011, led to the premature disengagement of the Bernstein centers and the United Kingdom Gatsby center(s) just before the start of the grant (Shtull-Trauring, 2012). An additional destructuring event occurred three years later, when the cognitive neuroscience pillar, directed by Stanislas Dehaene, was expelled from the main research axis by the triumvirate executive board coordinated by Henry Markram (an event recounted in Destexhe, 2021). This last move crystallized a deeper crisis between HBP and the neuroscience community (see the neurofuture letter coordinated by Mainen and Pouget, 2014; for review see Frégnac and Laurent, 2014; Mahfoud, 2021).

Did the flagship change the collective way we do brain sciences? Attracting and sustaining collaborations has been at the heart of the open calls of HBP (Lorents et al., 2023). Despite these efforts, monitored by ethnographic studies in social sciences (Mahfoud, 2018, 2021; Aicardi and Mahfoud, 2022; Rüland, 2023), it remains unclear at the end of the project what type of community has emerged or been consolidated. To my knowledge, no quantitative independent social network study has been yet made, looking at the full longitudinal history of the HBP flagship, to analyze (1) the evolution of collaborative networks (to illustrate for instance a possible revitalization by open calls) and (2) the interpenetration of the different scientific fields. Some of these networks, formed before HBP, had already been driving blue-sky projects (FACETS and BrainScaleS). These medium-sized interdisciplinary consortia became diluted in their merging with the partners of the Blue Brain Project (BBP; for review, see Frégnac and Laurent, 2014). The identification of fixed points (happy contributors constantly funded throughout the time course of HBP) is also missing. Did the funding mainly benefit preexisting scientific clusters, or sustain, from start to finish, a rich club of individual scientists? Did a new community of users of HBP infrastructures emerge in fine? My own impression is that the fixed points were tightly linked with the governance (EPFL, then Jülich) and that, around these, scientific subnetworks waxed and waned in succession. The consortium dynamics do not yet seem to have reached an equilibrium point where a stable envelope of labs could be constituted by regular users of HBP databases and digital infrastructures. This remains an open and fascinating issue for further social studies.

Building consortium coordination

A more down-to-earth related issue is to assess how the coordination of the scientific communities targeted by flagships operates, and what mechanisms can be used to measure its effectiveness. In the case of Graphene, the coordination was led by the Department of Physics at the Chalmers University in Sweden. A decade later, it is clear that this large-scale initiative has globally succeeded in its endeavors. According to the final report, scientific and technological promises have been fulfilled. Sixty to seventy per cent of what was promised in terms of specific applications has been delivered. A quantified study led by WifOR (commissioned by the flagship) makes the projection that Graphene will have created a total contribution to GDP of €3800 million and 38,400 new jobs in the 27 EU countries between 2014 and 2030. Per euro invested and compared with other EU projects, this flagship has performed 13 times better than expected in terms of patent applications, and seven times better for scientific publications. Seventeen spin-off companies have received over €130 million in private funding (Albu and Juneja, 2023; see also “Flagship EC reports” given underneath.).

• The European Commission (DG Connect) and the GRAPHENE and the HUMAN BRAIN PROJECT FET Flagships (2016). FET Flagships. Lessons learned from the first 30 months of their operation.https://www.h2020.md/en/fet-flagships-lessons-learntfet_flagships-lessons_learnedfromramp-upphase-2016oct 25final_18843.pdf

• See also Public presentation by Wolfgang Boch, FET-Flagship information day, Bratislava, SK, 23 May 2013: “FET Flagship Initiatives: Concept, Call and Evaluation results”

• The European Commission (DG Connect) and the GRAPHENE and the HUMAN BRAIN PROJECT FET Flagships (2023). Ten years of the European Union’s roll of the dice: €1B or 1billion euros each for the Human Brain Project (HBP) and the Graphene Flagship. 09 September 2023.https://www.frogheart.ca/?tag=graphene-week-september-5-9-2022-is-a-celebration-of-10-years-of-thegraphene-flagshiphttps://graphene-flagship.eu/materials/news/chalmers-steers-europe-s-major-graphene-venture-to-success/

Related statistics and economic studies are still missing concerning HBP, and it is likely that the Brain flagship did not provide a similar return of investment, compared with its Carbon flagship companion.

Indeed, the Graphene project met all the expected evaluation criteria, mostly because it was purely technology-driven and primarily engineered and organized as an industrial project. Things operated differently in the HBP, because of its initial blue-sky ambition (see above), and also in its coordination. The EC chose to run the flagship through the prestigious Ecole Polytechnnique Fédérale de Lausanne (EPFL), led by the ambitious entrepreneur Patrick Aebischer. To put it frankly, at the time no other institution of similar experience volunteered in Europe to run the HBP flagship. The choice of EPFL was justified by the fact that two of the members of the scientific triumvirate leading HBP, the charismatic Henry Markram and Richard Frackowiak, were professors in this respected institution. However, in addition to the fact that EPFL belonged to a non-EC country, the risk existed of a possible conflict of interest, since the EPFL was already engaged in the management of the Blue Brain Project (BBP) with a strong United States industrial partner, IBM. The initial choice of a scientific writer as a general communication manager could also indicate that more attention was given in the ramp-up phase to the control of the wording of what could be done, rather than to the consolidation of the scientific network while keeping the objectives within reasonable reach. Multiscale, multiomic digital reconstruction of the brain became the flagged mottos. Hype was there, from the start.

In terms of coordinating administration and the validated assessment of achievements, HBP created an administrative stranglehold. Rather than capitalizing on the local administrative resources already present in the participating European institutions, an oversized centralizing administration effort was constructed ahead of the consolidation of any scientific project. Fifty-two high-salary administrative positions were filled at the start of HBP. Lessons still remain to be learnt. The same issue is still present today into the management of EBRAINS, in part with the same players. An overpowering administrative approach was repeated again at the start of EBRAINS, when several dozens of such positions were requested (unsuccessfully) in the framework of the Horizon-Infra-2022-SERV call. Despite 10 years of experience, a good managerial model is yet to be found.

An objective evaluation of the intrinsic weight of the management budget, together with indirect coordinating activities (relayed by the partnering institutions), is difficult to reconstitute. In HBP the official administrative share of the total funding allowance was initially estimated to be 7–8%, i.e., 70–80 million euros (Markram, 2012). This projected administrative share was equal to the global budget of the Theoretical Neuroscience pillar! On top of that, indirect costs and overheads financed by HBP meant that administrative funding operated as a multilayer skimming of the initial endowment, dependent on the national institutions: each participating institution’s administration engulfed a sizeable share of what was supposed to be scientific funding. The real numbers do not seem to be retrievable in a transparent way, and, strangely, Chat-GPT breaks down if you ask the question to Alice or its deep learning Ersatz!!

A third issue has been how to monitor the effectiveness of coordination. Taken globally, the HBP flagship productivity remains difficult to evaluate: many objectives were renormalized during the course of the project, and these were reviewed for the EC by a multiplicity of experts, who, themselves, were often renewed at each phase of the contract. These different factors contributed to the absence of a common evaluation grid, which would have allowed a coherent tracking overview, from start to finish. This lack of longitudinal continuity finds an echo in the fact that many of the scientists who participated in the early phase of the Human Brain Project, including Henry Markram, the founder of HBP, were not present at its closing ceremonies (the Summit meeting in Marseille in March 2023, open to the public, and the more introvert closing event, held in Jülich in September 2023).

Concerning the scientific job market in HBP, unexpected trends were observed. Because of the relative lack of trust which appeared after the mediation between the EC and the various individual state administrations, the deliverable reporting and the level of surveillance of the productivity of each postdoctoral researcher, engineer, and even animal care technicians, were increased. The famous Key Performance Indicator (KPI) metrics, which fit with the managerial monitoring of industrialized and financial projects (Twin, 2023), were probably not best adapted to track progress in the more fundamental aspects of HBP. This holds in particular for highly diverse brain research studies, where the validation process must take into account exploratory experimentation and where publication of results usually takes several years. In parallel, tracing follow-through continuity was aggravated by employment rules at certain national administrations, for example the universities and the Centre National de la Recherche Scientifique (CNRS) in France: to avoid risking the obligation to transform project-specific, short-term recruitment into permanent tenured positions, individual postdoctoral researcher employment was limited to a two-year span (in fear of the application of the French Sauvadet Law). A possible solution might have been found by allowing rotations of the hiring institutions, so that the national constraints, linked with a fixed employer, could be overcome. Opening the possibility that a researcher or technician working at a given site of the European consortium could be financed in turn by different administrations would have fitted well with the European dimension of the project. Despite the crucial need to secure funding of key staff members for up to 10 years, the surprising overall reality of the HBP job market was that employment contracts were on average shorter than in regular research grants, or even FET grants. The unproductive consequence was a loss of continuity in the work dependent on successive postdocs and high-level engineers. A reduction in scientific efficacy toward reaching the “hard science” objectives and a down-calibration in the ambitions of the deliverables resulted from this fragmentation in hiring opportunities.

Building novel datasets

In terms of databases, considerable efforts in defining ontologies, including exhaustive analysis of metadata, are necessary ahead of any sharing of experimental work. It is only after agreeing on common classification labels that labs can interchange and convert their data around a common referential (Petilla Interneuron Nomenclature Group, 2008). Impressive attempts have been made beyond HBP, in particular by the Allen Institute (Koch and Reid, 2012), but large-scale initiatives have often kept their agenda and classification criteria separate, making a direct comparison difficult (Frégnac, 2017; Fairhall, 2021). In the case of HBP, no significant funding was allocated for the full implementation of interoperability of preexisting databases, including revisiting experimental lab notebooks according to accepted standards. This may have a cost in the long run: important structural and in vivo functional data, already acquired in species which have become out of fashion (cat, electric fish, giraffe.), might remain noncurated on some lab shelves. This difficult issue (no one wants to go back to old data formats) was already discussed between in vivo and in vitro neuroscientists at the time of exploratory projects of the FET initiative (Daisy, FACETS, BrainScaleS) in collaboration with the experts from the Blue Brain Project (BBP) at EPFL. Some aspects were partially explored by the INCF community (mostly neuroimaging and calcium imaging), leading to the widespread adoption of the Neurodata Without Borders format for new data acquisition and a solid semantic metadata framework (developed in the second half of HBP). It is however apparent at the beginning of EBRAINS (i.e., some 10 years later) that an unsatisfactory state remains, concerning mostly the functional levels of investigation where contextual metadata are impossible to reconstitute post hoc. For instance, still in 2023, the comparisons between functional in vivo and in vitro multiscale data remain inconclusive or superficial. It is fair to recognize that Henry Markram thought of this, and tried to annex INCF at the start of HBP, without much success at that time.

In terms of animal experimentation, HBP achievements have been limited by the contradiction between defining and implementing unified experimental paradigms, on the one hand, and using old data recorded in vastly different contexts, on the other hand. In principle, paradigmatic unification requires new experimentation sets. In reality, HBP limited the amount of animal experimentation in higher mammals (the species the closest to humans), focusing instead on rodents and tried to remove nonhuman primate (NHP) experimentation from its objectives, mostly for political correctness. Although this exclusion rule was relaxed in the second part of the grant, it greatly weakened interspecies comparisons and led to overreaching transfer of coding concepts, developed in rodents, to the human brain (Hodge, 2019; Lowe, 2019; Loomba et al., 2022).

By developing centralized database infrastructures, one benefit, foreseen by Henry Markram (Markram, 2012), was that the aggregation and mining of diverse long-tail data, as well as conversion of numerous small data sources into big data, would improve knowledge about neuroscience-related disorders (Markram, 2012; Ferguson et al., 2014). However, the shift in emphasis, during HBP, from efforts in data acquisition (mostly fine-scale multispecies physiology) and ontologies to the development of centralized database infrastructure (involving mostly exclusively neuroinformatics and HPC) and human neuroimaging had, in my view, an additional side-effect. Some governmental institutions in Europe (including France!) suggested that enough data may already be available on the laboratory shelves, constituting a pile of “siloed” dormant sources that just need to be curated (for review, see Choudhury et al., 2014; Frégnac, 2017). It seemed indeed easier in terms of budget control by institutions to turn scientists into high-tech engineers seeking patterns in existing data rather than to continue to fund basic research which requires new experimentation and costly animal care facilities. At the end of the road, the infrastructures for exploiting the datasets may be operational (Amunts et al., 2016), but what fraction of Europe’s neuroscience data has been curated to populate the promised HBP databases?

Building a coherent scientific strategy

The major novelty introduced by HBP has been to promote the concept of “virtual experimentation” (Markram et al., 2011; Markram, 2012), already applied in medical studies of human organs and biomedicine [see VHP, the virtual human physiology project in FP7 (Hunter et al., 2013); for review, see Mulugeta et al., 2018]. The concept was generalized in HBP to “in silico” hardware platforms, where detailed neuromorphic simulations of biological experiments were interpreted by physicists as some kind of experimentation (see theoretical discussion in Hadky, 2019). Note that the motivation here was not completely scientifically driven, but provided an easy answer to societal doubts and legitimate concerns about animal vivisection. Indeed, switching research experimentation from in situ to virtual in silico experimentation certainly conforms to the three Rs of ethics guidelines favored by the EC (replacement, reduction, refinement; see Russell and Burch, 1959). However, in my view, giving scientific legitimacy to such a trend is here more hazardous than it seems, and its generalized practice could be misleading. Computer simulations and virtual experimentation are presented to the public as if they could replace animal experimentation. They do not.

HBP relates the concept of virtual experimentation to that of “predictive neuroinformatics,” i.e., the idea to predict new findings from previous ones without ever carrying out the experiments to support (or reject) these hypotheses. In essence, this leads to a body of data, some of which are real, others are predicted, and yet others that are predicted from predicted data etc. This strategy in HBP was meant to overcome a key issue in which, because of technical limitations, much needed experiments could not be conducted on all spatial/temporal scales. In the end, this strategy generated a conglomerate of fictive data, which may be related to the models from which they originate, but cannot be attributed to the biological system under study. The situation is different in physics, where theoretical models are strong enough to correctly predict experimental facts. This challenge may be one bridge too far in brain sciences.

Strictly speaking, “virtual” experiments are experiments that test new hypotheses on real biological data that were collected to test other hypotheses. Classically, the qualifier “virtual” is used to refer to data collected from existing datasets (for instance, from different labs) and collated into a unique data set rather than taken from a new set of animals (Peterson, 1995). This concept in HBP, and in deep learning models, has been extended to the read-out of sets of local internal “hidden” variables, which are the exclusive results of simulations for different contexts, without being systematically validated by the experimenter. In most cases, the global parameters of the model were initially defined by the specific context of the biological experiment. They were trained to fit the correlation matrix between recorded real data, i.e., external input, biological data recorded in sparse specific brain sites and the global output performance. The same model can thus be replayed for other classes of simulated inputs, and virtual responses can be predicted under various types of perturbations produced by modulatory-like or electrical-like artificial manipulations, or even following in silico lesions (Markram, 2014). This data-enriching approach has generated a lot of interest in the modeling field. It has also been applied with some success to psychology and neurolinguistics (Jain et al., 2023). In physics of materials, deep learning algorithms trained by extensive structure-property datasets have recently been optimized to predict the capability of generating new materials with targeted properties (Honda et al., 2021). This example of experiment-free strategy certainly represents important advances that will influence future research.

Complementary virtual experimentation strategies have been used in the second half of HBP course, adding an elaborate twist of sophistication. The initial concept of “The Virtual Brain” (TVB) dates from 2008 (Ritter et al., 2013; Sanz, 2013), well before HBP. By merging individual anatomy from brain imaging data with state-of-the-art mathematical modeling, its aim was to reduce complexity on the microscopic level aiming to reveal the macroscopic organization. The driving hypothesis is that a TVB model of a human patient’s brain activity generates sufficiently accurate EEG, MEG, BOLD, and SEEG signals to reduce the dynamic complexity by a million-fold through methods from statistical physics (Ghosh et al., 2008). This hybrid approach, already applied in human medicine, has been merged during HBP with the concept of “digital twin,” which originates from the engineering industry, especially aeronautics (Grieves, 2019). A “digital twin” is a type of personalized computational brain model, based on measured real-world data obtained from its real-life counterpart, i.e., the patient. This concept has been applied, with impressive success, by Viktor Jirsa and colleagues in predictive modeling of focal epilepsy (Jirsa et al., 2023). In this case, sensor data are recorded in the real world and then used to build a virtual model of the individual brain. However, by feeding the databases with virtual data simulated in a metaverse world, the scientific project tends to suffer from a kind of poor man’s logic (I cannot record for real, so I simulate). We are no longer studying the brain but the properties of its digital Ersatz, as if these were precisely identical at every level of description and function. Identical “twins”: can still one pretend?

In biological, neurologic, or psychological terms, we are very far from understanding the complexity of brain function: to what extent is the construction of a digital facsimile actually advancing knowledge of the human mind? Indeed, is there any reason to expect that digital models would, could or should achieve capabilities of function and adaptation (in the case of lesion/inactivation paradigms) identical to those of the real brain. What are the critical parameters in the virtual brain model that are necessary to substantiate the biological relevance of the predicted effects? Can we exclude the possibility that the global simulation of the whole brain simply provides a hyped-up robe of metadata, used to visualize more localized perturbation dynamics in a richer context? In my view, additional targeted in situ experiments are necessary to understand the real differences, and to learn from these before making sense of the alternative source of knowledge provided by the virtual experimentation paradigm. To what extent is it a valid funding strategy to encourage “virtualization” as a way to replace in situ experimentation?

Recently, the HBP consortium has answered to these criticisms by distancing its current objectives from the multiomic initial approach of the ramp-up phase of the project (Markram et al., 2011; Markram, 2012). In Section 6 of the white Zenodo paper generated at the end of the HBP flagship (Amunts, et al., 2023), the reformed consortium brings a clear semantic clarification: “We distinguish purpose-driven digital twins from the abstract idea of a full digital replica (or duplicate/copy) of the brain, the latter being the complete representation of all aspects of the brain at all levels. A full replica of the brain is neither achievable nor does it seem of clear practical use.” Note however, that HBP does not take the full responsibility for its repeated use of the “twin” metaphor in relation to a biological brain, whose journalistic media impact (the “twin” duplicate), will unfortunately prevail in the eye of the public.

Building a unified multiscale theory of the brain

The global scientific strategy did change radically during the HBP flagship. At the start, “reverse engineering,” advocated by the charismatic Henry Markram, was the driving justification for the industrialization of neuroscience, the multiomic collection of big data and the need for high performance computing (HPC) infrastructures. One may agree or not with the soundness of the working hypothesis (see, for instance, Pitra, 2013; Frégnac, 2017), but it had the merit of providing a roadmap. Although doubtful about the Lego-like reconstruction strategy envisioned by Henry Markram, many of us thought that such an extensive reductionist exploration phase might uncover unsuspected alleys necessary for building a more comprehensive model of the brain.

The reductionist phase (Markram, 2012) progressively waned, when the triumvirate reduced experimental funding to the benefit of HPC, neuromorphic technology and future database infrastructures (Amunts et al., 2016, 2019), as if the community had siloed from the start enough low-level biological information to crack the brain. In the second part of the flagship, mean field (or neural mass) approaches and human brain imaging became dominant. In my view, this progressive shift in focus reflects the mid-flagship restructuration in governance, proposed by the Marquardt (2015) report, rather than a fundamental change in the way of looking at the brain. Science historians might make a finer diagnosis. Still, in a more scientific context, the shift could be interpreted as a down-grading of the initial ambitions, signaling the defeat of the theoretical field’s ambition to provide a multiscale view of the brain.

Let us expand a bit further. The main pragmatic issue of the reductionist phase during HBP was to decide how deep into the microscopic dimensions should the experimentation go. Unfortunately, we do not know yet. As early as 1714, the philosopher and mathematician Gottfried Wilhelm Leibniz argued that “if one could enter the brain as one enters a mill, there would be only mechanical parts, but one would not be able to observe thoughts” (cited in Cobb, 2020). Jonas and Kording (2017) confirmed the intuition of the “brain-mill” metaphor of Leibniz, by showing that reverse engineering methods would fall short of producing meaningful understanding of neural and computer systems, regardless of the amount of data. From the start of HBP, as well as in concurrent global neuroscience initiatives, a continuous debate has been whether reverse engineering is just a time/energy-consuming path, or a dead-end strategy (Pitra, 2013). Time will tell.

Following the transfer of the scientific leadership from EPFL to the Forschungszentrum Jülich, the HBP consortium gave less attention to molecular/cellular diversity, concentrating on brain imaging and mean field (neural mass) mesoscopic approaches. The mesoscopic dimension takes the lead in the public eye, mostly because striking correlations have been established between brain imaging mapping and perceptual or behavioral output. Even if fMRI signals are of a complex origin, mixing vascular, glial, and local field components rather than neuronal activity itself (Sirotin and Das, 2009; Schulz et al., 2012), they seem to provide topological markers of a medley of global activity. As such, they are used to encode the read-out of the holistic brain. This fits with what we know already from a phrenology-like localizationist approach (but with what statistical confidence? See Eklund et al., 2016; Murphy, 2016). This mapping strategy has been generalized with impressive success, to build, on the basis of more abstract labels, cognitive and semantic atlases (Huth et al., 2016).

In terms of computational neuroscience and theory, the HBP consortium has thus progressively shifted its focus of attention to the mesoscopic scale, to the detriment of pursuing an in-depth analysis of multiscale integration. The mean field formalism is attractive and seems the best candidate, since it provides a way to interpret brainwave diffusion in the context of functional brain maps. Its use has been most successful in understanding motor planning, decisional processes and speech production, all represented topographically in dedicated cortical output domains. Mean field models are also used to probe the dynamics of excitability states and their predictions guide cortical lesion experiments in epileptic patients.

However important hurdles persist. The mean field, or neural mass, certainly has global simplifying virtues (Deco et al., 2008; Pinotsis et al., 2014), but it relies on unproven assumptions of stochasticity and Gaussian distributions at more microscopic scales. The supposed stochasticity assumed by mesoscopic studies in fact ignores the biological diversity seen at the more microscopic level. This diversity is an essential functional specificity of central nervous system biology. Each molecular or cellular subelement is a carrier of retrievable information and its contribution to the mass (or, for the sake of illustration, columnar) activity field cannot be treated as purely additive noise or variability around a mean, as in traditional physics.

Other levels of complexification come with the multiscale organization of the living brain. For instance, efficient connectivity (functional connectome measures derived from fMRI resting state studies) in the brain does not depend only on the density of supposed connections and anatomic interaction distances (Mill et al., 2017). At a more microscopic level, functional wiring is specific to cell types and membrane affinities, it reflects the heterogeneity of local synaptic correlations and, very importantly, it is a dynamic function which depends on the timing-dependent integration of temporally structured factors and the history of past activity of the network. Consequently, smoothed (averaged) thresholded interpretation of fMRI studies leads to an interpretative view disconnected from the picture of integrative diversity obtained from results gathered in functional synaptic studies (Smith et al., 2013).

The same danger of disconnection may occur between mesoscopic and macroscopic scales, when the global geometry of the brain is taken into account. Recent work by Pang and colleagues, modeling large-scale brain activity (Pang et al., 2023), suggests that the macroscopic geometry of the brain may exert a more fundamental and stronger constraint on temporal “para”-synchronization dynamics than does the complex inter-regional mesoscopic connectivity (corresponding to the effective functional connectome). Knowledge of the connectome is not enough (Bargmann and Marder, 2013). It does not tell us much about how mental processes are generated, nor how different areas synchronize on wider scales. The real scientific challenge, which, in my view, would justify the HBP flagship dimension, is to account for the functional organization of the brain across all scales, beyond what can be charted topologically in a homeomorphic fashion on the 3D-cortical envelope of the brain.

We are far from establishing a comprehensive physical theory of the human brain. We still have to integrate conflicting observations at different scales. The success of mesoscopic approaches in the physics of inanimate matter is well established (“more is the same” in Kadanoff, 2009), but may not be strictly applicable to living glial and neuronal entities. Philip Anderson in his creative essay (“more is different”) described how new concepts, not present in ordinary classical or quantum mechanics, can arise from the consideration of aggregates of large numbers of particles (Anderson, 1972). With present knowledge, the system we study is necessarily a too complex product of evolution to be faithfully described by reduced equations. All we can do is to extract partial caricatures. These “simpler” representations can be used and manipulated to produce insightful predictions in specific contexts, but they may have, by construction and assumptions, eradicated important sources of causality and information. A new type of physics, of animated matter has to be invented, to better account for the multiscale stratification of information distributed in the living brain.

Consequently, and despite the claim of some HBP leaders (Amunts et al., 2019), integrating the three scales (micro-meso-macro) in a comprehensive functional model of the brain remains a challenge unsolved by HBP. Ten years later, we have still not decided whether spikes alone can yield useful simulations. We do not yet know if more microscopic variables and silent synaptic events, such as shunting inhibition, need to be considered, or whether dendro-dendritic computation may become effective in some globally ignited “conscious” states (Aru et al., 2023). Similarly, we suspect that more distributed or holistic information binding (glia-neuron communication, neuromodulation, nonsynaptic coupling effects; Cunha et al., 2022; Pinotsis and Miller, 2023) may be needed to account for the versatility and adaptivity of brain function. Perhaps now is the time to consider forming a more general quantitative model of brain networks not governed by statistical measures of association between spatial signals but based on the underlying most elemental physical properties of neural tissues from which those signals emerge (Van Horn et al., 2023).

The complexity of the model needed to accommodate the ever-growing flow of produced data may become commensurate with that of the brain itself (Borges, 1946)! In brief, simulating the full brain remains a target that is out of reach. Running simulations of its simplified caricature is certainly of major interest, but it should not make us forget that such models are not “twins” of the brain. As Steve Jobs, the founder of Apple, used to say: “Details matter. It’s worth waiting to get it right.”