Saturday, May 25, 2024
HomeArtificial IntelligenceWhat are Massive Language Fashions? What are they not?

What are Massive Language Fashions? What are they not?

“At this writing, the one severe ELIZA scripts which exist are some which trigger ELIZA to reply roughly as would sure psychotherapists (Rogerians). ELIZA performs finest when its human correspondent is initially instructed to”speak” to it, through the typewriter after all, simply as one would to a psychiatrist. This mode of dialog was chosen as a result of the psychiatric interview is without doubt one of the few examples of categorized dyadic pure language communication by which one of many taking part pair is free to imagine the pose of realizing nearly nothing of the true world. If, for instance, one have been to inform a psychiatrist “I went for a protracted boat trip” and he responded “Inform me about boats,” one wouldn’t assume that he knew nothing about boats, however that he had some objective in so directing the following dialog. You will need to observe that this assumption is one made by the speaker. Whether or not it’s practical or not is an altogether separate query. In any case, it has a vital psychological utility in that it serves the speaker to keep up his sense of being heard and understood. The speaker furher defends his impression (which even in actual life could also be illusory) by attributing to his conversational associate all kinds of background information, insights and reasoning means. However once more, these are the speaker’s contribution to the dialog.”

Joseph Weizenbaum, creator of ELIZA (Weizenbaum 1966).

GPT, the ancestor all numbered GPTs, was launched in June, 2018 – 5 years in the past, as I write this. 5 years: that’s a very long time. It definitely is as measured on the time scale of deep studying, the factor that’s, normally, behind when individuals speak of “AI.” One yr later, GPT was adopted by GPT-2; one other yr later, by GPT-3. At this level, public consideration was nonetheless modest – as anticipated, actually, for these sorts of applied sciences that require a lot of specialist information. (For GPT-2, what might have elevated consideration past the conventional, a bit, was OpenAI ’s refusal to publish the whole coaching code and full mannequin weights, supposedly because of the menace posed by the mannequin’s capabilities – alternatively, as argued by others, as a advertising and marketing technique, or but alternatively, as a technique to protect one’s personal aggressive benefit only a tiny little bit longer.

As of 2023, with GPT-3.5 and GPT-4 having adopted, the whole lot seems totally different. (Virtually) everybody appears to know GPT, no less than when that acronym seems prefixed by a sure syllable. Relying on who you speak to, individuals don’t appear to cease speaking about that incredible [insert thing here] ChatGPT generated for them, about its monumental usefulness with respect to [insert goal here]… or in regards to the flagrant errors it made, and the hazard that authorized regulation and political enforcement won’t ever be capable of catch up.

What made the distinction? Clearly, it’s ChatGPT, or put in another way, the truth that now, there’s a means for individuals to make lively use of such a device, using it for no matter their private wants or pursuits are. In truth, I’d argue it’s greater than that: ChatGPT is just not some impersonal device – it talks to you, selecting up your clarifications, adjustments of subject, temper… It’s somebody somewhat than one thing, or no less than that’s the way it appears. I’ll come again to that time in It’s us, actually: Anthropomorphism unleashed. Earlier than, let’s check out the underlying know-how.

Massive Language Fashions: What they’re

How is it even doable to construct a machine that talks to you? A technique is to have that machine hear rather a lot. And hear is what these machines do; they do it rather a lot. However listening alone would by no means be sufficient to realize outcomes as spectacular as these we see. As a substitute, LLMs observe some type of “maximally lively listening”: Repeatedly, they attempt to predict the speaker’s subsequent utterance. By “repeatedly,” I imply word-by-word: At every coaching step, the mannequin is requested to supply the following phrase in a textual content.

Possibly in my final sentence, you famous the time period “practice.” As per frequent sense, “coaching” implies some type of supervision. It additionally implies some type of technique. Since studying materials is scraped from the web, the true continuation is at all times recognized. The precondition for supervision is thus at all times fulfilled: A supervisor can simply examine mannequin prediction with what actually follows within the textual content. Stays the query of technique. That’s the place we have to speak about deep studying, and we’ll do this in Mannequin coaching.

Total structure

Right this moment’s LLMs are, in a roundabout way or the opposite, based mostly on an structure generally known as the Transformer. This structure was initially launched in a paper catchily titled “Consideration is all you want” (Vaswani et al. 2017). After all, this was not the primary try at automating natural-language era – not even in deep studying, the sub-type of machine studying whose defining attribute are many-layered (“deep”) synthetic neural networks. However there, in deep studying, it constituted some type of paradigm change. Earlier than, fashions designed to unravel sequence-prediction duties (time-series forecasting, textual content era…) tended to be based mostly on some type of recurrent structure, launched within the 1990’s (eternities in the past, on the time scale of deep-learning) by (Hochreiter and Schmidhuber 1997). Principally, the idea of recurrence, with its related threading of a latent state, was changed by “consideration.” That’s what the paper’s title was meant to speak: The authors didn’t introduce “consideration”; as an alternative, they essentially expanded its utilization in order to render recurrence superfluous.

How did that ancestral Transformer look? – One prototypical process in pure language processing is machine translation. In translation, be it finished by a machine or by a human, there may be an enter (in a single language) and an output (in one other). That enter, name it a code. Whoever needs to ascertain its counterpart within the goal language first must decode it. Certainly, one in all two top-level constructing blocks of the archetypal Transformer was a decoder, or somewhat, a stack of decoders utilized in succession. At its finish, out popped a phrase within the goal language. What, then, was the opposite high-level block? It was an encoder, one thing that takes textual content (or tokens, somewhat, i.e., one thing that has undergone tokenization) and converts it right into a kind the decoder could make sense of. (Clearly, there is no such thing as a analogue to this in human translation.)

From this two-stack structure, subsequent developments tended to maintain only one. The GPT household, along with many others, simply saved the decoder stack. Now, doesn’t the decoder want some type of enter – if to not translate to a special language, then to answer to, as within the chatbot situation? Seems that no, it doesn’t – and that’s why you may as well have the bot provoke the dialog. Unbeknownst to you, there’ll, in reality, be an enter to the mannequin – some type of token signifying “finish of enter.” In that case, the mannequin will draw on its coaching expertise to generate a phrase prone to begin out a phrase. That one phrase will then turn out to be the brand new enter to proceed from, and so forth. Summing up up to now, then, GPT-like LLMs are Transformer Decoders.

The query is, how does such a stack of decoders reach fulfilling the duty?

GPT-type fashions up shut

In opening the black field, we deal with its two interfaces – enter and output – in addition to on the internals, its core.


For simplicity, let me communicate of phrases, not tokens. Now think about a machine that’s to work with – extra even: “perceive” – phrases. For a pc to course of non-numeric information, a conversion to numbers essentially has to occur. The easy technique to effectuate that is to resolve on a hard and fast lexicon, and assign every phrase a quantity. And this works: The way in which deep neural networks are educated, they don’t want semantic relationships to exist between entities within the coaching information to memorize formal construction. Does this imply they may seem excellent whereas coaching, however fail in real-world prediction? – If the coaching information are consultant of how we converse, all can be wonderful. In a world of excellent surveillance, machines might exist which have internalized our each spoken phrase. Earlier than that occurs, although, the coaching information can be imperfect.

A way more promising method than to easily index phrases, then, is to signify them in a richer, higher-dimensional house, an embedding house. This concept, in style not simply in deep studying however in pure language processing general, actually goes far past something domain-specific – linguistic entities, say. You might be able to fruitfully make use of it in just about any area – offered you may devise a technique to sensibly map the given information into that house. In deep studying, these embeddings are obtained in a intelligent means: as a by-product of kinds of the general coaching workflow. Technically, that is achieved by way of a devoted neural-network layer tasked with evolving these mappings. Notice how, good although this technique could also be, it implies that the general setting – the whole lot from coaching information through mannequin structure to optimization algorithms employed – essentially impacts the ensuing embeddings. And since these could also be extracted and made use of in down-stream duties, this issues.

As to the GPT household, such an embedding layer constitutes a part of its enter interface – one “half,” so to say. Technically, the second makes use of the identical sort of layer, however with a special objective. To distinction the 2, let me spell out clearly what, within the half we’ve talked about already, is getting mapped to what. The mapping is between a phrase index – a sequence 1, 2, …, <vocabulary dimension> – on the one hand and a set of continuous-valued vectors of some size – 100, say – on the opposite. (One in all them might like this: (start{bmatrix} 1.002 & 0.71 & 0.0004 &… finish{bmatrix})) Thus, we receive an embedding for each phrase. However language is greater than an unordered meeting of phrases. Rearranging phrases, if syntactically allowed, might end in drastically modified semantics. Within the pre-transformer paradigma, threading a sequentially-updated hidden state took care of this. Put in another way, in that sort of mannequin, details about enter order by no means received misplaced all through the layers. Transformer-type architectures, nonetheless, must discover a totally different means. Right here, a wide range of rivaling strategies exists. Some assume an underlying periodicity in semanto-syntactic construction. Others – and the GPT household, as but and insofar we all know, has been a part of them – method the problem in precisely the identical means as for the lexical items: They make studying these so-called place embeddings a by-product of mannequin coaching. Implementation-wise, the one distinction is that now the enter to the mapping seems like this: 1, 2, …, <most place> the place “most place” displays alternative of maximal sequence size supported.

Summing up, verbal enter is thus encoded – embedded, enriched – twofold because it enters the machine. The 2 kinds of embedding are mixed and handed on to the mannequin core, the already-mentioned decoder stack.

Core Processing

The decoder stack is made up of some variety of similar blocks (12, within the case of GPT-2). (By “similar” I imply that the structure is similar; the weights – the place the place a neural-network layer shops what it “is aware of” – are usually not. Extra on these “weights” quickly.)

Inside every block, some sub-layers are just about “enterprise as common.” One is just not: the eye module, the “magic” ingredient that enabled Transformer-based architectures to forego retaining a latent state. To elucidate how this works, let’s take translation for instance.

Within the classical encoder-decoder setup, the one most intuitive for machine translation, think about the very first decoder within the stack of decoders. It receives as enter a length-seven cypher, the encoded model of an unique length-seven phrase. Since, resulting from how the encoder blocks are constructed, enter order is conserved, we now have a devoted illustration of source-language phrase order. Within the goal language, nonetheless, phrase order could be very totally different. A decoder module, in producing the interpretation, had somewhat not do that by translating every phrase because it seems. As a substitute, it will be fascinating for it to know which among the many already-seen tokens is most related proper now, to generate the very subsequent output token. Put in another way, it had higher know the place to direct its consideration.

Thus, determine the best way to distribute focus is what consideration modules do. How do they do it? They compute, for every accessible input-language token, how good a match, a match, it’s for their very own present enter. Keep in mind that each token, at each processing stage, is encoded as a vector of steady values. How good a match any of, say, three source-language vectors is is then computed by projecting one’s present enter vector onto every of the three. The nearer the vectors, the longer the projected vector. Based mostly on the projection onto every source-input token, that token is weighted, and the eye module passes on the aggregated assessments to the following neural-network module.

To elucidate what consideration modules are for, I’ve made use of the machine-translation situation, a situation that ought to lend a sure intuitiveness to the operation. However for GPT-family fashions, we have to summary this a bit. First, there is no such thing as a encoder stack, so “consideration” is computed amongst decoder-resident tokens solely. And second – keep in mind I stated a stack was constructed up of similar modules? – this occurs in each decoder block. That’s, when intermediate outcomes are bubbled up the stack, at every stage the enter is weighted as applicable at that stage. Whereas that is tougher to intuit than what occurred within the translation situation, I’d argue that within the summary, it makes quite a lot of sense. For an analogy, think about some type of hierarchical categorization of entities. As higher-level classes are constructed from lower-level ones, at every stage the method wants to have a look at its enter afresh, and resolve on a smart means of subsuming similar-in-some-way classes.


Stack of decoders traversed, the multi-dimensional codes that come out must be transformed into one thing that may be in contrast with the precise phrase continuation we see within the coaching corpus. Technically, this entails a projection operation as nicely a technique for choosing the output phrase – that phrase in target-language vocabulary that has the very best likelihood. How do you resolve on a technique? I’ll say extra about that within the part Mechanics of textual content era, the place I assume a chatbot consumer’s perspective.

Mannequin coaching

Earlier than we get there, only a fast phrase about mannequin coaching. LLMs are deep neural networks, and as such, they’re educated like all community is. First, assuming you’ve entry to the so-called “floor reality,” you may at all times examine mannequin prediction with the true goal. You then quantify the distinction – by which algorithm will have an effect on coaching outcomes. Then, you talk that distinction – the loss – to the community. It, in flip, goes by its modules, from again/high to start out/backside, and updates its saved “information” – matrices of steady numbers referred to as weights. Since data is handed from layer to layer, in a course reverse to that adopted in computing predictions, this method is named back-propagation.

And all that’s not triggered as soon as, however iteratively, for a sure variety of so-called “epochs,” and modulated by a set of so-called “hyper-parameters.” In observe, quite a lot of experimentation goes into deciding on the best-working configuration of those settings.

Mechanics of textual content era

We already know that in mannequin coaching, predictions are generated word-by-word; at each step, the mannequin’s information about what has been stated up to now is augmented by one token: the phrase that basically was following at that time. If, making use of a educated mannequin, a bot is requested to answer to a query, its response should by necessity be generated in the identical means. Nevertheless, the precise “right phrase” is just not recognized. The one means, then, is to feed again to the mannequin its personal most up-to-date prediction. (By necessity, this lends to textual content era a really particular character, the place each choice the bot makes co-determines its future conduct.)

Why, although, speak about selections? Doesn’t the bot simply act on behalf of the core mannequin, the LLM – thus passing on the ultimate output? Not fairly. At every prediction step, the mannequin yields a vector, with values as many as there are entries within the vocabulary. As per mannequin design and coaching rationale, these vectors are “scores” – scores, form of, how good a match a phrase can be on this scenario. Like in life, increased is healthier. However that doesn’t imply you’d simply choose the phrase with the very best worth. In any case, these scores are transformed to possibilities, and an acceptable likelihood distribution is used to non-deterministically choose a probable (or likely-ish) phrase. The likelihood distribution generally used is the multinomial distribution, applicable for discrete alternative amongst greater than two options. However what in regards to the conversion to possibilities? Right here, there may be room for experimentation.

Technically, the algorithm employed is named the softmax perform. It’s a simplified model of the Boltzmann distribution, well-known in statistical mechanics, used to acquire the likelihood of a system’s state provided that state’s power and the temperature of the system. However for temperature, each formulae are, in reality, similar. In bodily techniques, temperature modulates possibilities within the following means: The warmer the system, the nearer the states’ possibilities are to one another; the colder it will get, the extra distinct these possibilities. Within the excessive, at very low temperatures there can be a couple of clear “winners” and a silent majority of “losers.”

In deep studying, a like impact is straightforward to realize (by way of a scaling issue). That’s why you’ll have heard individuals speak about some bizarre factor referred to as “temperature” that resulted in [insert adjective here] solutions. If the applying you employ helps you to differ that issue, you’ll see {that a} low temperature will end in deterministic-looking, repetitive, “boring” continuations, whereas a excessive one might make the machine seem as if it have been on medicine.

That concludes our high-level overview of LLMs. Having seen the machine dissected on this means might have already got left you with some form of opinion of what these fashions are – not. This subject greater than deserves a devoted exposition – and papers are being written pointing to essential facets on a regular basis – however on this textual content, I’d prefer to no less than supply some enter for thought.

Massive Language Fashions: What they don’t seem to be

Partially one,describing LLMs technically, I’ve generally felt tempted to make use of phrases like “understanding” or “information” when utilized to the machine. I could have ended up utilizing them; in that case, I’ve tried to recollect to at all times encompass them with quotes. The latter, the including quotes, stands in distinction to many texts, even ones revealed in an educational context (Bender and Koller 2020). The query is, although: Why did I even really feel compelled to make use of these phrases, given I do not assume they apply, of their common that means? I can consider a easy – shockingly easy, possibly – reply: It’s as a result of us, people, we predict, speak, share our ideas in these phrases. After I say perceive, I surmise you’ll know what I imply.

Now, why do I feel that these machines don’t perceive human language, within the sense we normally suggest when utilizing that phrase?

A couple of info

I’ll begin out briefly mentioning empirical outcomes, conclusive thought experiments, and theoretical concerns. All facets touched upon (and lots of extra) are greater than worthy of in-depth dialogue, however such dialogue is clearly out of scope for this synoptic-in-character textual content.

First, whereas it’s arduous to place a quantity on the standard of a chatbot’s solutions, efficiency on standardized benchmarks is the “bread and butter” of machine studying – its reporting being a necessary a part of the prototypical deep-learning publication. (You can even name it the “cookie,” the driving incentive, since fashions normally are explicitly educated and fine-tuned for good outcomes on these benchmarks.) And such benchmarks exist for a lot of the down-stream duties the LLMs are used for: machine translation, producing summaries, textual content classification, and even somewhat ambitious-sounding setups related to – quote/unquote – reasoning.

How do you assess such a functionality? Right here is an instance from a benchmark named “Argument Reasoning Comprehension Job” (Habernal et al. 2018).

Declare: Google is just not a dangerous monopoly
Purpose: Folks can select to not use Google
Warrant: Different engines like google don’t redirect to Google
Various: All different engines like google redirect to Google

Right here declare and motive collectively make up the argument. However what, precisely, is it that hyperlinks them? At first look, this may even be complicated to a human. The lacking hyperlink is what known as warrant right here – add it in, and all of it begins to make sense. The duty, then, is to resolve which of warrant or various helps the conclusion, and which one doesn’t.

If you consider it, this can be a surprisingly difficult process. Particularly, it appears to inescapingly require world information. So if language fashions, as has been claimed, carry out almost in addition to people, it appears they should have such information – no quotes added. Nevertheless, in response to such claims, analysis has been carried out to uncover the hidden mechanism that allows such seemingly-superior outcomes. For that benchmark, it has been discovered (Niven and Kao 2019) that there have been spurious statistical cues in the best way the dataset was constructed – these eliminated, LLM efficiency was no higher than random.

World information, in reality, is without doubt one of the fundamental issues an LLM lacks. Bender et al. (Bender and Koller 2020) convincingly reveal its essentiality by way of two thought experiments. One in all them, located on a lone island, imagines an octopus inserting itself into some cable-mediated human communication, studying the chit-chat, and at last – having gotten bored – impersonating one of many people. This works wonderful, till someday, its communication associate finds themselves in an emergency, and must construct some rescue device out of issues given within the atmosphere. They urgently ask for recommendation – and the octopus has no concept what to reply. It has no concepts what these phrases truly discuss with.

The opposite argument comes straight from machine studying, and strikingly easy although it might be, it makes its level very nicely. Think about an LLM educated as common, together with on a lot of textual content involving vegetation. It has additionally been educated on a dataset of unlabeled images, the precise process being unsubstantial – say it needed to fill out masked areas. Now, we pull out an image and ask: What number of of that blackberry’s blossoms have already opened? The mannequin has no likelihood to reply the query.

Now, please look again on the Joseph Weizenbaum quote I opened this text with. It’s nonetheless true that language-generating machine haven’t any information of the world we stay in.

Earlier than transferring on, I’d like to only shortly trace at a completely totally different sort of consideration, introduced up in a (2003!) paper by Spärck Jones (Spaerck 2004). Although written lengthy earlier than LLMs, and lengthy earlier than deep studying began its profitable conquest, on an summary degree it’s nonetheless very relevant to at this time’s scenario. Right this moment, LLMs are employed to “study language,” i.e., for language acquisition. That talent is then constructed upon by specialised fashions, of task-dependent structure. Fashionable real-world down-stream duties are translation, doc retrieval, or textual content summarization. When the paper was written, there was no such two-stage pipeline. The creator was questioning the match between how language modeling was conceptualized – particularly, as a type of restoration – and the character of those down-stream duties. Was restoration – inferring a lacking, for no matter causes – piece of textual content a very good mannequin, of, say, condensing a protracted, detailed piece of textual content into a brief, concise, factual one? If not, might the rationale it nonetheless appeared to work simply wonderful be of a really totally different nature – a technical, operational, coincidental one?

[…] the essential characterisation of the connection between the enter and the output is in reality offloaded within the LM method onto the selection of coaching information. We are able to use LM for summarising as a result of we all know that some set of coaching information consists of full texts paired with their summaries.

It appears to me that at this time’s two-stage course of however, that is nonetheless a side price giving some thought.

It’s us: Language studying, shared objectives, and a shared world

We’ve already talked about world information. What else are LLMs lacking out on?

In our world, you’ll hardly discover something that doesn’t contain different individuals. This goes rather a lot deeper than the simply observable info: our continuously speaking, studying and typing messages, documenting our lives on social networks… We don’t expertise, discover, clarify a world of our personal. As a substitute, all these actions are inter-subjectively constructed. Emotions are. Cognition is; that means is. And it goes deeper but. Implicit assumptions information us to continuously search for that means, be it in overheard fragments, mysterious symbols, or life occasions.

How does this relate to LLMs? For one, they’re islands of their very own. While you ask them for recommendation – to develop a analysis speculation and an identical operationalization, say, or whether or not a detainee needs to be launched on parole – they haven’t any stakes within the final result, no motivation (be it intrinsic or extrinsic), no objectives. If an harmless particular person is harmed, they don’t really feel the regret; if an experiment is profitable however lacks explanatory energy, they don’t sense the vanity; if the world blows up, it gained’t have been their world.

Secondly, it’s us who’re not islands. In Bender et al.’s octopus situation, the human on one facet of the cable performs an lively position not simply after they communicate. In making sense of what the octopus says, they contribute a necessary ingredient: particularly, what they assume the octopus needs, thinks, feels, expects… Anticipating, they replicate on what the octopus anticipates.

As Bender et al. put it:

It’s not that O’s utterances make sense, however somewhat, that A could make sense of them.

That article (Bender and Koller 2020) additionally brings spectacular proof from human language acquisition: Our predisposition in the direction of language studying however, infants don’t study from the supply of enter alone. A scenario of joint consideration is required for them to study. Psychologizing, one might hypothesize they should get the impression that these sounds, these phrases, and the actual fact they’re linked collectively, truly issues.

Let me conclude, then, with my ultimate “psychologization.”

It’s us, actually: Anthropomorphism unleashed

Sure, it’s wonderful what these machines do. (And that makes them extremely harmful energy devices.) However this by no means impacts the human-machine variations which were present all through historical past, and live on at this time. That we’re inclined to assume they perceive, know, imply – that possibly even they’re acutely aware: that’s on us. We are able to expertise deep feelings watching a film; hope that if we simply attempt sufficient, we are able to sense what a distant-in-evolutionary-genealogy creature is feeling; see a cloud encouragingly smiling at us; learn an indication in an association of pebbles.

Our inclination to anthropomorphize is a present; however it might generally be dangerous. And nothing of that is particular to the twenty-first century.

Like I started with him, let me conclude with Weizenbaum.

Some topics have been very arduous to persuade that ELIZA (with its current script) is not human.

Picture by Marjan
on Unsplash

Bender, Emily M., and Alexander Koller. 2020. “Climbing In the direction of NLU: On Which means, Type, and Understanding within the Age of Knowledge.” In Proceedings of the 58th Annual Assembly of the Affiliation for Computational Linguistics, 5185–98. On-line: Affiliation for Computational Linguistics.
Caliskan, Aylin, Pimparkar Parth Ajay, Tessa Charlesworth, Robert Wolfe, and Mahzarin R. Banaji. 2022. “Gender Bias in Phrase Embeddings.” In Proceedings of the 2022 AAAI/ACM Convention on AI, Ethics, and Society. ACM.
Habernal, Ivan, Henning Wachsmuth, Iryna Gurevych, and Benno Stein. 2018. “The Argument Reasoning Comprehension Job: Identification and Reconstruction of Implicit Warrants.” In Proceedings of the 2018 Convention of the North American Chapter of the Affiliation for Computational Linguistics: Human Language Applied sciences, Quantity 1 (Lengthy Papers), 1930–40. New Orleans, Louisiana: Affiliation for Computational Linguistics.
Hochreiter, Sepp, and Jürgen Schmidhuber. 1997. “Lengthy Quick-Time period Reminiscence.” Neural Computation 9 (December): 1735–80.
Niven, Timothy, and Hung-Yu Kao. 2019. “Probing Neural Community Comprehension of Pure Language Arguments.” CoRR abs/1907.07355.

Spaerck, Karen. 2004. “Language Modelling’s Generative Mannequin : Is It Rational?” In.

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. “Consideration Is All You Want.”
Weizenbaum, Joseph. 1966. “ELIZA – a Pc Program for the Examine of Pure Language Communication Between Man and Machine.” Commun. ACM 9 (1): 36–45.



Please enter your comment!
Please enter your name here

Most Popular

Recent Comments