Despite the preposterous premise, it's one of the best, most enjoyable books I've ever read. I got mine for a song (an unsold hardback in nearly mint condition) through Amazon's second hand book service.
Ooh, testy. Is this one of those irregular verbs? - you PROVIDE facts; - they DISTORT facts; - we (right thinking people) have CONCENSUS; - they (oil company shills) are DELUDED or FRAUDULENT.
Of course, the tone of your post suggests that your position is neither ideological or authoritarian.
As I scientist I see it as my job to be skeptical - that is, to pay at least as much attention to falsification as to confirmation. I'm constantly amazed at the shrillness of the climate debate where, for most, scientific detachment seems to have been replaced with emotional involvement.
Singularity runs as a collection of Software Isolated Processes (SIPs) that (normally) run in a single address space. Each SIP appears to the kernel as a thread. SIPs can have multiple threads. SIPs can provide their own memory management and thread management. SIPs communicate through a shared-memory message passing interface where precisely one SIP has ownership of a shared memory block at any given time.
Another late reply - sorry! This is exactly the problem. You could fairly define AI to be "all the computational problems we have no idea how to tackle using existing science and engineering". A large part of any AI PhD is defining precisely what the problem is that you want to tackle. By the time you've done that (no mean feat), you realise that barring a flash of genius you will only be able to take baby steps towards a solution.
As to why I got into AI: it was because I read way too much science fiction:-) I still think it's a worthy goal, but these days I suspect the most fruitful route lies in building computational models of psychological theories rather than the more traditional theorem-prover/neural-net-in-a-box approach. Of course, the psychologists are probably doing a great deal of hand waving in their own theories.
A late reply, but, IMHO, the problem with "Society of Mind" (and other books in that vein) is that it's basically just an appealing metaphor rather than a concrete guide. You read it, you think you have a handle on this idea of "mind", then find yourself floundering when you sit down to write an AI program based on the ideas in the book. This is what happened to me and, judging by the complete lack of success elsewhere, I assume it's what's happened to everybody else. I even asked Minsky at a conference whether he'd written any programs based on his "Society of Mind" idea; he responded that, being a mathematician, he didn't do implementation work.
Speaking as someone with a PhD in AI, I'm very, very skeptical about having human-level AI by 2029.
Whatever definition of intelligence you choose, it probably includes learning and reasoning components. We have some effective learning algorithms, provided your domain is very specific and you have boat loads of training data. We have next to no good reasoning algorithms. Complete search is a dead duck and incomplete search is not very reliable. Worse, search algorithms get seriously confused when the data base is inconsistent (humans are good at maintaining several incompatible world models simultaneously). And that's all before you consider that we have no psychological models of human reasoning that are anywhere near being specific enough to guide an implementation project (please don't mention "Society of Mind"). Finally, there is precious little funding out there for this kind of research, which is a shame, but there you go.
Unfortunately the majority of people obtaining a CS degree get stuck on some particular tool (e.g., Java) and spend the rest of their career suffering from tunnel vision (every problem looks like a nail). Your goal should be to learn enough of the basics and theory that you can pick up the textbook for any given technology and be up and running within a week or two. Here's a list off the top of my head of things you should understand:
[Hardware] The fetch/decode/execute cycle Instruction decoding Pipelining, out-of-order execution, etc. Caches TLBs Busses Interrupts Discs Networking hardware basics
[Software] Programming language paradigms: imperative, functional, OO, logical Finite automata Grammars and parsing Data structures: lists, queues (FIFO and priority), dictionaries, sets, trees, hash tables Algorithms: sorting, searching, etc. Memory management and garbage collection Big O notation Type systems (in particular, Hindley-Milner style) and type checking/inference Basics of security and encryption The relational model for databases Concurrency: pi calculus (or some relation), locking and lock management, software transactional memory Distributed systems: data exchange, error handling, establishing communication channels Queueing theory basics: Markov chains etc. Programming language implementation basics
No doubt I've missed a bunch of important things off the list.
All of the mundane stuff about learning various programming languages and toolkits should be seen as a means to explore the above topics, not as an end in itself.
Many thanks, I appreciate you taking the time to craft a long response.
Unfortunately I find many of the explanations you don't completely convince me (which is not to say you are wrong or that I cannot be convinced). Here are some problems: (1) why is the antarctic expanding (other than a minor calving peninsula)? Surely the same processes affecting the arctic are also at work in the antarctic. (2) What are the error bars on the proxies you describe? Could you point me to any references explaining (a) the physical processes connecting the proxies to ice thickness and (b) the experimental data validating the hypothesis? (3) I follow your explanation of what happens in summer, but I don't follow how this affects the ice that grows in winter. (4) How do we know that the ice that recently melted did not previously melt in summer, say, 1000 years ago?
The real question, as far as I'm concerned, is not whether the globe is warming - I'm pretty sure it is - but whether and how much of it is anthropogenic or natural. I have been following this debate fairly closely the last few years and was quite surprised at how flakey the AGW argument is, which is why I remain skeptical. For what it's worth, I am a practicing scientist. While I'm not a climate scientist, I flatter myself that I can follow a well presented argument and I can spot the difference between a hand-waving explanation and a solid piece of science. I've spent some time at RealClimate.org, but frankly it's far too one-sided to be considered a real scientific forum. Not to mention the RC prediliction for censoring awkward questions. Have you read climateaudit.org? If not, you should take a look. CA makes it clear the AGW argument is not even close to settled for anyone requiring standard scientific rigour.
But doesn't IIA just mean that the voting scheme should produce the same overall relative preference between A and B regardless of whether we consider voter preferences for C or not? That is, the voters can make their decisions however they like, but the overall preference between A and B is decided purely on the basis of the voters' preferences for A w.r.t. B (after taking transitivity into account). This seems entirely reasonable in my opinion.
I'm curious. The satellite data only goes back about thirty years, just about enough for a single datum in climate terms. It seems quite a leap to suggest this is unprecedented (my apologies if I have misunderstood you) from the data available. Can you give me a reference to the historical data on ice thickness? What is your evidence for blaming global warming on thinning ice? I thought ice accumulation was due to precipitation, which is supposed to go up with the temperature, not down, according to my understanding of AGW theory.
Something else you might be able to explain for me: looking at this NASA time-lapse film of a year in the life of the arctic, the annual variance in ice coverage is enormous - on the order of six to seven million square kilometres of the stuff appears and disappears every year. How does accumulated warming affect ice that isn't there for nearly half the year?
I'm asking in good faith. I'm not out for a fight.
Regarding the epithet "denialist", can you name any individual who actually deserves it?
That's very kind of you, but I am quite capable of reading the graph. For one thing, the graph shows global ice coverage, not just the arctic. Is global warming not happening in the antarctic? For another thing, if you think you can see a meaningful trend in four years of data that noisy then I have some homeopathic remedies to sell you. Finally, the global ice coverage in 1998 (to pick one of the years you mention) was considerably lower than it is at present, going by the graph.
From memory, Arrow's Impossibility Theorem applies to voting systems satisfying the following criteria: (1) an individual vote is a consistent (i.e., acyclic) set of preferences between candidates (i.e., a possibly constrained set of "I prefer A to B" statements); (2) the result of the election is a consistent set of preferences between candidates; (3) [majority rule] if all voters prefer A to B then the result must prefer A to B; (4) [independence] whether or not C is a candidate should not be able to prejudice the relative preference between A and B. Arrow's theorem shows that the only voting scheme satisfying these criteria is a dictatorship (i.e., only one distinguished voter's preferences count).
These criteria seem entirely reasonable to me. While range voting is not subject to Arrow's theorem (voters assign scores to candidates rather than preferences between candidates), it does not respect criterion (3), which I find rather unpalatable.
Sorry, isn't ice cover currently upa million square kilometres (globally) compared to the average? Of course, as a skeptic (I beg your pardon, "denialist"), I'm far too stupid and evil for anything I say to have value.
That's interesting. This paper on the aftermath of Chernobyl on Finland, which was in the zone of greatest fallout, indicates no increase in the incidence of childhood leukaemia.
Ahh, the classic RealClimate debating technique: attack the messenger, not the message. (I haven't even looked at the web site, but I really don't care *who* says something, I only care about *what* they say.)
Says McKitrick. When Mann et al. removed one or even several proxies from the reconstruction, they still got the hockey stick. And, as I may remind you again, the NRC and NAS both found that the hockey stick is a robust feature of the data even when they did not agree with Mann's methods.
Mann et al did not remove the BCPs which were the only proxies carrying the hockeystick signal. McKitrick demonstrated that (along with showing that the MBH technique produced a hockeystick 99% of the time even when fed red noise...) The NAS report merely concluded that the MBH result was "plausible", hardly resounding support for the work.
I wondered about error bars on the temperature data, you responded by pointing out the "error bars" on the model predictions. On the contrary, you merely stated that climate models "do not calculate error bars". Of course they do not calculate error bars on data; they are models. They do calculate error bars on their predictions. Instrumental records and proxy reconstructions are what calculate error bars on data.
No, you are conflating two things that I said: (1) the graph to which you referred does not show error bars for the mean temperature as measured (this has nothing to do with the models); (2) GCMs do not compute error bars on their predictions - they don't and, again, I don't accept accept that comparing an "ensemble" of runs tells you anything more than the distribution of results of the models. Put another way, the models are discrete approximations of physical processes and therefore are bound to introduce error in the results. What is this error? We don't know because it isn't computed.
Everything in AR4 is based on literature published in or before 2006. It does not produce new research, it only summarizes the existing literature. The relevant literature is there for your reading pleasure. You can either look for it yourself, or wait for the other AR4 chapters to be published which cite that literature.
You can't begin by pointing me at a summary document as a reference with some authority and finish by telling me to go read everything on which it might be based. Either the summary is authoritative or it is not, make up your mind.
It's worthy of your flat claim, which was backed up by no facts or tests of statistical significance. But to amplify, GCMs have many parameters but their projections are dominated by relatively few of these, as demonstrated by sensitivity analysis. And, as I have also noted before, for better understanding of the probability distribution you can turn to EMICs, which are computationally simpler but are rapidly becoming competitive with GCMs for practical purposes.
It takes weeks or months to run a model. Just how many of these can you get through with a given model? How large is the parameter space?
The point is that (a) changing the time at which the data is read does cancel out if it is merely random deviation from the standard time (assuming it's not like a 12 hour difference or something), and that (b) if it is a consistent day-to-day bias which makes a significant difference to the data instead of a random deviation, it will likely show up as an outlier and be discarded from the analysis as untrustworthy, unless (c) you are prepared to claim that a substantial number of stations all made substantial time errors all in the same reinforcing direction.
(a) I can easily imagine reading times changing from dawn to noon to dusk at different stations; (b) how do you know when you're discarding an "outlier" versus, say, having an airport built nearby? (c) I have no idea about the size of bias, if any, but then nor do you. Which brings me back to my original question about the accuracy of instrument readings taken by humans across the globe over the last century or so. Assuming all or most of them to be
Wait a minute. If you remove the bristlecone pines from Mann's reconstruction then the hockeystick disappears. The whole point Wegman (and M&M) were making is that Mann's method essentially data-mines for hockeystick shapes. The NAS panel even concluded that BCPs were not suitable for use in climate reconstructions, not that that has prevented recent reconstructions from reusing them (which, in turn, raises questions of the statistical validity of "independent" reconstructions that use the same data over and over again - a process which surely violates the principle of random sampling).
As for your "refutations", most of them are simply flat contradictions of what I said.
Let's see:
- I wondered about error bars on the temperature data, you responded by pointing out the "error bars" on the model predictions. Not the same thing.
- I questioned the utility of a summary document made by and for policy makers and pointed out that the actual assessment report has, amazingly, not even been released. You said I should go and read the literature, but the point is that the relevant literature (AR4) is not available.
- I question your claim that just because GCM parameters can't be tuned to produce *any* result that doesn't mean that the models haven't been overfit to the data (overfitting here meaning minimising the error w.r.t. the training data at the cost of increasing the error w.r.t. the testing data). You simply repeated what you said in the first place. I don't care whether a model has zero parameters or a hundred, what I care about is its predictive power. (This ties into my point about error bars on the temperature data.)
- I claim that doing a relatively small number of GCM runs on a small sample of the possible parameter settings is not statistically significant (the number of possible parameter settings is enormous, the number of experiments you can do in a reasonable amount of time is very small). Your response was just a flat denial.
- I said that taking the average of an ensemble of different GCM results does not tell you anything about the error of any given GCM, it only tells you something about the distribution of results from the GCMs, not the quality of the results. You said "More nonsense" (one of those lovely debating techniques that makes discussions with the blind faithful such a pleasure) and went on about a totally different point comparing GCM results against reconstructions.
- I wondered about the accuracy of temperature readings taken over the last century or so. You seemed to suggest that the time of day at which they are read makes little difference (surely I must have misunderstood you here?) and claimed that the errors pretty much all cancel out. The point is that changing the time at which a reading occurs introduces a systematic error into the data.
- You state that dendroproxy data contains useful information, without justification, and that dendro reconstructions don't use the linearity assumption. As I understand it, that is exactly the assumption used in such reconstructions. If you can point me to a document correcting me I would be very grateful.
- I said that if you take out CO2 forcing from the models then global warming disappears entirely. You said you have to take out all the other GHGs too. But, according to Hansen at least, the contribution of the other major GHGs is almost exactly balanced by cooling anthropogenic forcings from aerosols and so forth. So your refutation does not stand.
- I suggested that Jones' UHI results were questionable (at the very least they are unreproducible since he has so far refused to disclose the precise methods and data he used) and, even though Jones' paper is still being referenced, you suggested I look at something more recent. All right, how about Streutker 2003 (just google for "UHI Houston"). To quote from co2science.org:
in just twelve years the UHI of Houston grew by more than the IPCC contends the mean surface air temperature
More nonsense. Spoken like someone who gets all his climate science from ClimateAudit.
Spoken like someone who gets all his science from RealClimate. I cannot take you seriously when you suggest that the NAS report supported the hockey stick. But that aside, what about the Wegman report? Did you read it? Did you understand it? Wegman (chair of the National Academy of Sciences' Committee on Applied and Theoretical Statistics) demolished the statistical foundations of the hockey stick and entirely supported the work of McIntyre and McKitrick.
I'm rather dubious of the value of the SPM - a summary by policy makers for policy makers (and why on earth wasn't AR4 released before the summary?). For example, figure SPM-4 suggests global land temperatures have risen 1C over the 20th century. That is substantially larger than the other estimates I've seen. Moreover, there are no error bars on the global land temperature line. Is this the upper bound? The lower bound? How were these data computed? In short, it's not a scientifically useful figure.
Regarding parameter tuning, the whole reason for having the parameters is because your model doesn't work without them. Having complete parameterisability has nothing to do with overfitting; you can overfit any parameterised model to the training data.
On that note, comparing multiple runs of multiple models emphatically does not give you error bars. It merely allows you to compare the outputs of models for a statistically insignificant sample of their parameter settings. It says nothing whatsoever about the validity of those outputs.
On the UHI, I do not believe it has been shown to be negligible. The most widely cited paper on the topic, Jones et al. 1990, provides neither the data nor the method used and, astoundingly, Jones refuses to provide this information when asked (see, e.g., www.climateaudit.org ad nauseam). Even more interestingly, two of that paper's coauthors published another UHI paper the same year, but with completely different results and even remarked that "The differences over the 1954-83 period indicate that since the late 1970s the rate of warming at urban stations is over 0.1 deg C decade relative to more rural stations.... These results suggest that caution must be used when using trends from stations in the vicinity of major metropolitan centers. [...] Our work differs from the recent study by Jones et al 1990. They have shown that any urban bias in their data has been mitigated over eastern China. The reasons for this are not clear." Plus there's the perhaps minor point that Jones considers a local population of 150,000 as rural.
I am not at all surprised that models do well compared to weather station measurements, because they are trained on that data! My point was that the weather station measurements themselves contain large sources of error. Even if you ignore the possibility of the UHI effect, which I don't, I do not believe that thermometers read by humans are going to give more than 1C of accuracy or that they have been consistently read at the same times each day for a century or that most of them haven't been moved around. Plus there is the fact that there are not very many of them and virtually all of them are in well inhabited areas. That surely should set some alarm bells ringing.
As far as proxy records are concerned, are you seriously trying to defend things like dendroclimatology? The hockey stick has been utterly debunked. You will be hard pressed to find a dendrologist who will agree with the proposition that temperature is the key driver of tree ring widths and, moreover, that ring widths are linearly related to temperature. Other proxies such as bulloides etc. are also widely taken as proxies for other factors (winds, currents, etc.) Anyone who thinks they are going to extract a definite temperature signal with an error of less than 1C is surely kidding themselves.
When I say that GCMs essentially assume CO2 to be the only forcing agent, what I mean is that if you remove the calculated contribution from CO2 from the models, you get virtually no global warming. Do you not find this suspicious?
Regarding estimates of population/economic growth, you say that the GCMs, which you are defending, still use out of date models. Those models are vastly inaccurate, in the sense that they predict five or six billion more people by the end of the 21st century and predict that the average income in, say, Brazil, will significantly outstrip that of the USA.
You can call my criticisms outdated, but as far as I can tell they remain unaddressed. Perhaps you should take a more honestly skeptical position.
(a) disagree with one another by a factor of three,
(b) even their least pessimistic predictions have been substantially in excess of the observed temperature rise,
(c) depend upon the fine tuning of a large number of parameters (i.e., fudge factors to account for either failings of the model or as abstractions of physical processes that are not modelled, such as cloud formation),
(d) do not calculate error bars,
(e) are calibrated against a temperature record that is fraught with error (e.g.: weather stations largely cover only the relatively small area of well inhabited areas of the globe; there is the urban heat island effect; there is disagreement between satellite, balloon,and weather staion records; temperature records do not go back very far; results obtained via proxies such as tree rings are very questionable and hard to connect to the instrumental record),
(f) essentially assume that CO2 is the *only* climate forcing agent,
(g) assume vastly more pessimistic population and economic future growth than professional demographers and economists support.
Despite the amount of money and effort put into them, there is every reason to be very skeptical of the output of computer climate models.
Slashdot Mirror
You're not a climate scientist, are you? :-)
Despite the preposterous premise, it's one of the best, most enjoyable books I've ever read. I got mine for a song (an unsold hardback in nearly mint condition) through Amazon's second hand book service.
Now that's what I call a compelling argument!
Ooh, testy. Is this one of those irregular verbs?
- you PROVIDE facts;
- they DISTORT facts;
- we (right thinking people) have CONCENSUS;
- they (oil company shills) are DELUDED or FRAUDULENT.
Of course, the tone of your post suggests that your position is neither ideological or authoritarian.
I hate to burst your bubble, but even ignoring the high point of 1998, the trend over the last ten years is cooling if you don't cherry pick your source (see e.g., http://rankexploits.com/musings/2008/ipcc-projections-continue-to-falsify/ which uses GISS, HadCRU, MSU, and NOAA).
As I scientist I see it as my job to be skeptical - that is, to pay at least as much attention to falsification as to confirmation. I'm constantly amazed at the shrillness of the climate debate where, for most, scientific detachment seems to have been replaced with emotional involvement.
Singularity runs as a collection of Software Isolated Processes (SIPs) that (normally) run in a single address space. Each SIP appears to the kernel as a thread. SIPs can have multiple threads. SIPs can provide their own memory management and thread management. SIPs communicate through a shared-memory message passing interface where precisely one SIP has ownership of a shared memory block at any given time.
Another late reply - sorry! This is exactly the problem. You could fairly define AI to be "all the computational problems we have no idea how to tackle using existing science and engineering". A large part of any AI PhD is defining precisely what the problem is that you want to tackle. By the time you've done that (no mean feat), you realise that barring a flash of genius you will only be able to take baby steps towards a solution.
:-) I still think it's a worthy goal, but these days I suspect the most fruitful route lies in building computational models of psychological theories rather than the more traditional theorem-prover/neural-net-in-a-box approach. Of course, the psychologists are probably doing a great deal of hand waving in their own theories.
As to why I got into AI: it was because I read way too much science fiction
A late reply, but, IMHO, the problem with "Society of Mind" (and other books in that vein) is that it's basically just an appealing metaphor rather than a concrete guide. You read it, you think you have a handle on this idea of "mind", then find yourself floundering when you sit down to write an AI program based on the ideas in the book. This is what happened to me and, judging by the complete lack of success elsewhere, I assume it's what's happened to everybody else. I even asked Minsky at a conference whether he'd written any programs based on his "Society of Mind" idea; he responded that, being a mathematician, he didn't do implementation work.
Speaking as someone with a PhD in AI, I'm very, very skeptical about having human-level AI by 2029.
Whatever definition of intelligence you choose, it probably includes learning and reasoning components. We have some effective learning algorithms, provided your domain is very specific and you have boat loads of training data. We have next to no good reasoning algorithms. Complete search is a dead duck and incomplete search is not very reliable. Worse, search algorithms get seriously confused when the data base is inconsistent (humans are good at maintaining several incompatible world models simultaneously). And that's all before you consider that we have no psychological models of human reasoning that are anywhere near being specific enough to guide an implementation project (please don't mention "Society of Mind"). Finally, there is precious little funding out there for this kind of research, which is a shame, but there you go.
Unfortunately the majority of people obtaining a CS degree get stuck on some particular tool (e.g., Java) and spend the rest of their career suffering from tunnel vision (every problem looks like a nail). Your goal should be to learn enough of the basics and theory that you can pick up the textbook for any given technology and be up and running within a week or two. Here's a list off the top of my head of things you should understand:
[Hardware]
The fetch/decode/execute cycle
Instruction decoding
Pipelining, out-of-order execution, etc.
Caches
TLBs
Busses
Interrupts
Discs
Networking hardware basics
[Software]
Programming language paradigms: imperative, functional, OO, logical
Finite automata
Grammars and parsing
Data structures: lists, queues (FIFO and priority), dictionaries, sets, trees, hash tables
Algorithms: sorting, searching, etc.
Memory management and garbage collection
Big O notation
Type systems (in particular, Hindley-Milner style) and type checking/inference
Basics of security and encryption
The relational model for databases
Concurrency: pi calculus (or some relation), locking and lock management, software transactional memory
Distributed systems: data exchange, error handling, establishing communication channels
Queueing theory basics: Markov chains etc.
Programming language implementation basics
No doubt I've missed a bunch of important things off the list.
All of the mundane stuff about learning various programming languages and toolkits should be seen as a means to explore the above topics, not as an end in itself.
Hope this helps,
-- Ralph
Many thanks, I appreciate you taking the time to craft a long response.
Unfortunately I find many of the explanations you don't completely convince me (which is not to say you are wrong or that I cannot be convinced). Here are some problems:
(1) why is the antarctic expanding (other than a minor calving peninsula)? Surely the same processes affecting the arctic are also at work in the antarctic.
(2) What are the error bars on the proxies you describe? Could you point me to any references explaining (a) the physical processes connecting the proxies to ice thickness and (b) the experimental data validating the hypothesis?
(3) I follow your explanation of what happens in summer, but I don't follow how this affects the ice that grows in winter.
(4) How do we know that the ice that recently melted did not previously melt in summer, say, 1000 years ago?
The real question, as far as I'm concerned, is not whether the globe is warming - I'm pretty sure it is - but whether and how much of it is anthropogenic or natural. I have been following this debate fairly closely the last few years and was quite surprised at how flakey the AGW argument is, which is why I remain skeptical. For what it's worth, I am a practicing scientist. While I'm not a climate scientist, I flatter myself that I can follow a well presented argument and I can spot the difference between a hand-waving explanation and a solid piece of science. I've spent some time at RealClimate.org, but frankly it's far too one-sided to be considered a real scientific forum. Not to mention the RC prediliction for censoring awkward questions. Have you read climateaudit.org? If not, you should take a look. CA makes it clear the AGW argument is not even close to settled for anyone requiring standard scientific rigour.
-- Ralph
But doesn't IIA just mean that the voting scheme should produce the same overall relative preference between A and B regardless of whether we consider voter preferences for C or not? That is, the voters can make their decisions however they like, but the overall preference between A and B is decided purely on the basis of the voters' preferences for A w.r.t. B (after taking transitivity into account). This seems entirely reasonable in my opinion.
I'm curious. The satellite data only goes back about thirty years, just about enough for a single datum in climate terms. It seems quite a leap to suggest this is unprecedented (my apologies if I have misunderstood you) from the data available. Can you give me a reference to the historical data on ice thickness? What is your evidence for blaming global warming on thinning ice? I thought ice accumulation was due to precipitation, which is supposed to go up with the temperature, not down, according to my understanding of AGW theory.
Something else you might be able to explain for me: looking at this NASA time-lapse film of a year in the life of the arctic, the annual variance in ice coverage is enormous - on the order of six to seven million square kilometres of the stuff appears and disappears every year. How does accumulated warming affect ice that isn't there for nearly half the year?
I'm asking in good faith. I'm not out for a fight.
Regarding the epithet "denialist", can you name any individual who actually deserves it?
That's very kind of you, but I am quite capable of reading the graph. For one thing, the graph shows global ice coverage, not just the arctic. Is global warming not happening in the antarctic? For another thing, if you think you can see a meaningful trend in four years of data that noisy then I have some homeopathic remedies to sell you. Finally, the global ice coverage in 1998 (to pick one of the years you mention) was considerably lower than it is at present, going by the graph.
From memory, Arrow's Impossibility Theorem applies to voting systems satisfying the following criteria:
(1) an individual vote is a consistent (i.e., acyclic) set of preferences between candidates (i.e., a possibly constrained set of "I prefer A to B" statements);
(2) the result of the election is a consistent set of preferences between candidates;
(3) [majority rule] if all voters prefer A to B then the result must prefer A to B;
(4) [independence] whether or not C is a candidate should not be able to prejudice the relative preference between A and B.
Arrow's theorem shows that the only voting scheme satisfying these criteria is a dictatorship (i.e., only one distinguished voter's preferences count).
These criteria seem entirely reasonable to me. While range voting is not subject to Arrow's theorem (voters assign scores to candidates rather than preferences between candidates), it does not respect criterion (3), which I find rather unpalatable.
Sorry, isn't ice cover currently up a million square kilometres (globally) compared to the average? Of course, as a skeptic (I beg your pardon, "denialist"), I'm far too stupid and evil for anything I say to have value.
That's interesting. This paper on the aftermath of Chernobyl on Finland, which was in the zone of greatest fallout, indicates no increase in the incidence of childhood leukaemia.
That's the funniest thing I've read in ages.
The author was joking, right?
Ahh, the classic RealClimate debating technique: attack the messenger, not the message. (I haven't even looked at the web site, but I really don't care *who* says something, I only care about *what* they say.)
Mann et al did not remove the BCPs which were the only proxies carrying the hockeystick signal. McKitrick demonstrated that (along with showing that the MBH technique produced a hockeystick 99% of the time even when fed red noise...) The NAS report merely concluded that the MBH result was "plausible", hardly resounding support for the work.
No, you are conflating two things that I said: (1) the graph to which you referred does not show error bars for the mean temperature as measured (this has nothing to do with the models); (2) GCMs do not compute error bars on their predictions - they don't and, again, I don't accept accept that comparing an "ensemble" of runs tells you anything more than the distribution of results of the models. Put another way, the models are discrete approximations of physical processes and therefore are bound to introduce error in the results. What is this error? We don't know because it isn't computed.
You can't begin by pointing me at a summary document as a reference with some authority and finish by telling me to go read everything on which it might be based. Either the summary is authoritative or it is not, make up your mind.
It takes weeks or months to run a model. Just how many of these can you get through with a given model? How large is the parameter space?
(a) I can easily imagine reading times changing from dawn to noon to dusk at different stations;
(b) how do you know when you're discarding an "outlier" versus, say, having an airport built nearby?
(c) I have no idea about the size of bias, if any, but then nor do you. Which brings me back to my original question about the accuracy of instrument readings taken by humans across the globe over the last century or so. Assuming all or most of them to be
As for your "refutations", most of them are simply flat contradictions of what I said.
Let's see:
- I wondered about error bars on the temperature data, you responded by pointing out the "error bars" on the model predictions. Not the same thing.
- I questioned the utility of a summary document made by and for policy makers and pointed out that the actual assessment report has, amazingly, not even been released. You said I should go and read the literature, but the point is that the relevant literature (AR4) is not available.
- I question your claim that just because GCM parameters can't be tuned to produce *any* result that doesn't mean that the models haven't been overfit to the data (overfitting here meaning minimising the error w.r.t. the training data at the cost of increasing the error w.r.t. the testing data). You simply repeated what you said in the first place. I don't care whether a model has zero parameters or a hundred, what I care about is its predictive power. (This ties into my point about error bars on the temperature data.)
- I claim that doing a relatively small number of GCM runs on a small sample of the possible parameter settings is not statistically significant (the number of possible parameter settings is enormous, the number of experiments you can do in a reasonable amount of time is very small). Your response was just a flat denial.
- I said that taking the average of an ensemble of different GCM results does not tell you anything about the error of any given GCM, it only tells you something about the distribution of results from the GCMs, not the quality of the results. You said "More nonsense" (one of those lovely debating techniques that makes discussions with the blind faithful such a pleasure) and went on about a totally different point comparing GCM results against reconstructions.
- I wondered about the accuracy of temperature readings taken over the last century or so. You seemed to suggest that the time of day at which they are read makes little difference (surely I must have misunderstood you here?) and claimed that the errors pretty much all cancel out. The point is that changing the time at which a reading occurs introduces a systematic error into the data.
- You state that dendroproxy data contains useful information, without justification, and that dendro reconstructions don't use the linearity assumption. As I understand it, that is exactly the assumption used in such reconstructions. If you can point me to a document correcting me I would be very grateful.
- I said that if you take out CO2 forcing from the models then global warming disappears entirely. You said you have to take out all the other GHGs too. But, according to Hansen at least, the contribution of the other major GHGs is almost exactly balanced by cooling anthropogenic forcings from aerosols and so forth. So your refutation does not stand.
- I suggested that Jones' UHI results were questionable (at the very least they are unreproducible since he has so far refused to disclose the precise methods and data he used) and, even though Jones' paper is still being referenced, you suggested I look at something more recent. All right, how about Streutker 2003 (just google for "UHI Houston"). To quote from co2science.org:
Spoken like someone who gets all his science from RealClimate. I cannot take you seriously when you suggest that the NAS report supported the hockey stick. But that aside, what about the Wegman report? Did you read it? Did you understand it? Wegman (chair of the National Academy of Sciences' Committee on Applied and Theoretical Statistics) demolished the statistical foundations of the hockey stick and entirely supported the work of McIntyre and McKitrick.
I'm rather dubious of the value of the SPM - a summary by policy makers for policy makers (and why on earth wasn't AR4 released before the summary?). For example, figure SPM-4 suggests global land temperatures have risen 1C over the 20th century. That is substantially larger than the other estimates I've seen. Moreover, there are no error bars on the global land temperature line. Is this the upper bound? The lower bound? How were these data computed? In short, it's not a scientifically useful figure.
... These results suggest that caution must be used when using trends from stations in the vicinity of major metropolitan centers. [...] Our work differs from the recent study by Jones et al 1990. They have shown that any urban bias in their data has been mitigated over eastern China. The reasons for this are not clear." Plus there's the perhaps minor point that Jones considers a local population of 150,000 as rural.
Regarding parameter tuning, the whole reason for having the parameters is because your model doesn't work without them. Having complete parameterisability has nothing to do with overfitting; you can overfit any parameterised model to the training data.
On that note, comparing multiple runs of multiple models emphatically does not give you error bars. It merely allows you to compare the outputs of models for a statistically insignificant sample of their parameter settings. It says nothing whatsoever about the validity of those outputs.
On the UHI, I do not believe it has been shown to be negligible. The most widely cited paper on the topic, Jones et al. 1990, provides neither the data nor the method used and, astoundingly, Jones refuses to provide this information when asked (see, e.g., www.climateaudit.org ad nauseam). Even more interestingly, two of that paper's coauthors published another UHI paper the same year, but with completely different results and even remarked that "The differences over the 1954-83 period indicate that since the late 1970s the rate of warming at urban stations is over 0.1 deg C decade relative to more rural stations.
I am not at all surprised that models do well compared to weather station measurements, because they are trained on that data! My point was that the weather station measurements themselves contain large sources of error. Even if you ignore the possibility of the UHI effect, which I don't, I do not believe that thermometers read by humans are going to give more than 1C of accuracy or that they have been consistently read at the same times each day for a century or that most of them haven't been moved around. Plus there is the fact that there are not very many of them and virtually all of them are in well inhabited areas. That surely should set some alarm bells ringing.
As far as proxy records are concerned, are you seriously trying to defend things like dendroclimatology? The hockey stick has been utterly debunked. You will be hard pressed to find a dendrologist who will agree with the proposition that temperature is the key driver of tree ring widths and, moreover, that ring widths are linearly related to temperature. Other proxies such as bulloides etc. are also widely taken as proxies for other factors (winds, currents, etc.) Anyone who thinks they are going to extract a definite temperature signal with an error of less than 1C is surely kidding themselves.
When I say that GCMs essentially assume CO2 to be the only forcing agent, what I mean is that if you remove the calculated contribution from CO2 from the models, you get virtually no global warming. Do you not find this suspicious?
Regarding estimates of population/economic growth, you say that the GCMs, which you are defending, still use out of date models. Those models are vastly inaccurate, in the sense that they predict five or six billion more people by the end of the 21st century and predict that the average income in, say, Brazil, will significantly outstrip that of the USA.
You can call my criticisms outdated, but as far as I can tell they remain unaddressed. Perhaps you should take a more honestly skeptical position.
Unfortunately the models
(a) disagree with one another by a factor of three,
(b) even their least pessimistic predictions have been substantially in excess of the observed temperature rise,
(c) depend upon the fine tuning of a large number of parameters (i.e., fudge factors to account for either failings of the model or as abstractions of physical processes that are not modelled, such as cloud formation),
(d) do not calculate error bars,
(e) are calibrated against a temperature record that is fraught with error (e.g.: weather stations largely cover only the relatively small area of well inhabited areas of the globe; there is the urban heat island effect; there is disagreement between satellite, balloon,and weather staion records; temperature records do not go back very far; results obtained via proxies such as tree rings are very questionable and hard to connect to the instrumental record),
(f) essentially assume that CO2 is the *only* climate forcing agent,
(g) assume vastly more pessimistic population and economic future growth than professional demographers and economists support.
Despite the amount of money and effort put into them, there is every reason to be very skeptical of the output of computer climate models.
The science at the Real Climate site is often called into question. For balance, readers should also take a look at, say, Climate Audit.