SoylentNews

On that last point, we have some interesting features. First, a graph titled "Forecast evaluation for models run in 2004" which doesn't mention a source or what adjustments were made. Second, the final bit makes a claim that 17 models were evaluated for accuracy. What they neglected to tell you is that the actual number is far greater than 17. In the actual research [wiley.com], look at figure 2 on page 5. Note the labels "IPCC FAR, IPCC SAR, and IPCC TAR". Those refer to assessment reports by the IPCC (FAR - First Assessment Report, etc). Those are aggregations of dozens of models each not single models. (Elsewhere in the report was a mention of "AR4" which is the fourth Assessment Report, but it doesn't show up in this particular figure.) From page 3:

Starting in the mid‐1990s, climate modeling efforts were primarily undertaken in conjunction with the IPCC process (and later, the Coupled Model Intercomparison Projects, CMIPs), and model projections were taken from models featured in the IPCC FAR (1990), Second Assessment Report (SAR‐IPCC, 1996), Third Assessment Report (TAR‐IPCC, 2001), and Fourth Assessment Report (AR4‐IPCC, 2007).

So of the 17 model projections, the last four or so are actually massive aggregations of models. And the report found problems with the IPCC aggregates. On page 5:

However, the remaining eight models—RS71, H81 Scenario 1, H88 Scenarios A, B, and C, FAR,MS93, and TAR—had projected forcings significantly stronger or weaker than observed (Figure 1).

For IPCC FAR and IPCC TAR (note they are referred to as "models"), it was stronger, a typical symptom of "running hot" bias. Notice also the use of an "implied TCR metric" throughout the paper to ignore these differences.

We compared observations to climate model projections over the model projection period using two approaches: change in temperature versus time and change in temperature versus change in radiative forcing (“implied TCR”). We use an implied TCR metric to provide a meaningful model‐observation comparison even in the presence of forcing differences. Implied TCR is calculated by regressing temperature change against radiative forcing for both models and observations, and multiplying the resulting values by the forcing associated with doubled atmospheric CO2concentrations, F2x, (following Otto et al., 2013):

A typical policy question is when you emit a certain amount of CO2 equivalent, what global temperature increase do you see? By going to a purely radiative forcing viewpoint, they ignore both the increased removal of greenhouse gases from the atmosphere and the positive feedback mechanisms that are supposed to result in higher temperatures in the long term. But that's a huge part of the policy decision!

To be blunt, your link was significantly dishonest. It alleged to compare 17 "models", but we find that it's actually comparing hundreds with at least two large aggregates showing the consistent bias problems I discussed earlier, and emphasizes metrics that hide those bias issues.