Content Form:APRJA 13 Pierre Depaz - Revision history

Simoon at 14:16, 5 November 2024

2024-11-05T14:16:24Z

← Older revision		Revision as of 14:16, 5 November 2024
Line 125:		Line 125:
	== Notes ==		== Notes ==

	<div class="~~1column~~">		<div class="notes">

	This article has benefited greatly from thorough discussions with, and copy edits by, Sara Messelaar Hammerschmidt.		This article has benefited greatly from thorough discussions with, and copy edits by, Sara Messelaar Hammerschmidt.

			</div>

			<div class="1column">

	<references />		<references />

Simoon at 14:12, 5 November 2024

2024-11-05T14:12:49Z

← Older revision		Revision as of 14:12, 5 November 2024
Line 124:		Line 124:

	== Notes ==		== Notes ==

			<div class="1column">

	This article has benefited greatly from thorough discussions with, and copy edits by, Sara Messelaar Hammerschmidt.		This article has benefited greatly from thorough discussions with, and copy edits by, Sara Messelaar Hammerschmidt.

	~~<div class="1column">~~

	<references />		<references />

Geoffcox at 10:01, 15 October 2024

2024-10-15T10:01:24Z

← Older revision		Revision as of 10:01, 15 October 2024
Line 23:		Line 23:
	Discussing Alan Turing’s proposal of machine intelligence as an educational problem, we conclude by turning to theories of co-construction of intelligence (Bachimont; Stiegler) to sketch out, through examples of linguistic normalization, hallucinations, and prompting, how such word embeddings can operate logics of control themselves.		Discussing Alan Turing’s proposal of machine intelligence as an educational problem, we conclude by turning to theories of co-construction of intelligence (Bachimont; Stiegler) to sketch out, through examples of linguistic normalization, hallucinations, and prompting, how such word embeddings can operate logics of control themselves.

	== From a ~~bit~~ to a ~~vector~~ ==		== From a Bit to a Vector ==
	The question of discursive communication in technical systems is inseparable from the question of encoding. Whether as frequency-modulated hertzian waves, pixel arrays, or smoke clouds, different encodings enable different discourses (Postman). This section focuses on the shift from one encoding to the other and its semantic implications, looking at both the bit and the vector as a means to represent information in digital environments and highlighting how sense-making shifts from one to the other.		The question of discursive communication in technical systems is inseparable from the question of encoding. Whether as frequency-modulated hertzian waves, pixel arrays, or smoke clouds, different encodings enable different discourses (Postman). This section focuses on the shift from one encoding to the other and its semantic implications, looking at both the bit and the vector as a means to represent information in digital environments and highlighting how sense-making shifts from one to the other.

Line 53:		Line 53:
	Reconstituting meaning from binary encoding has always been a clearly defined problem, involving only mathematical reconstitution of the original message. Correctness of meaning, on the other hand, began as a computer-syntactic problem, but shifted with vectors to become a human-semantic problem.		Reconstituting meaning from binary encoding has always been a clearly defined problem, involving only mathematical reconstitution of the original message. Correctness of meaning, on the other hand, began as a computer-syntactic problem, but shifted with vectors to become a human-semantic problem.

	== Shaping ~~vectors~~ ==		== Shaping Vectors ==
	We now turn our attention to techniques deployed by producers of LLMs to shape word embeddings of LLMs into models capable of meaningful output. After looking at the use of benchmarks for capability discovery, we argue that these processes operate as a form of discipline, as theorized by Michel Foucault. Then, we turn to reinforcement learning as an example of such shaping, but this time through the lens of a form of control, following Gilles Deleuze. We then conclude this section by reframing ''training'' in terms of ''education'', drawing on Alan Turing’s seminal paper, “Computing Machinery and Intelligence” (1956).		We now turn our attention to techniques deployed by producers of LLMs to shape word embeddings of LLMs into models capable of meaningful output. After looking at the use of benchmarks for capability discovery, we argue that these processes operate as a form of discipline, as theorized by Michel Foucault. Then, we turn to reinforcement learning as an example of such shaping, but this time through the lens of a form of control, following Gilles Deleuze. We then conclude this section by reframing ''training'' in terms of ''education'', drawing on Alan Turing’s seminal paper, “Computing Machinery and Intelligence” (1956).

Line 76:		Line 76:
	Such modulation happens through processes known as reinforcement learning, whether with human or AI feedback. Reinforcement learning judges each output against standards to support subsequent optimization. It involves having a trusted authority (such as a human who has been told what to expect from an ideal LLM output) provide feedback to the training model to reinforce certain semantic features (e.g., preventing any output that is deemed discriminatory, copyright infringement, or harmful to the user) (Kaelbling).		Such modulation happens through processes known as reinforcement learning, whether with human or AI feedback. Reinforcement learning judges each output against standards to support subsequent optimization. It involves having a trusted authority (such as a human who has been told what to expect from an ideal LLM output) provide feedback to the training model to reinforce certain semantic features (e.g., preventing any output that is deemed discriminatory, copyright infringement, or harmful to the user) (Kaelbling).

	While benchmarking focused on abstract comparability through normative testing, reinforcement learning involves more subjective normalization of meaning through feedback and iteration in order to align the model with what is considered a legally, morally, and socially acceptable meaning.		While benchmarking focused on abstract comparability through normative testing, reinforcement learning involves more subjective normalization of meaning through feedback and iteration in order to align the model with what is considered a legally, morally, and socially acceptable meaning. Such logic updates a disciplinary approach to undetermined behaviour and enters the realm of control. In his 1992 essay, Gilles Deleuze describes a new kind of era, ushered by a new kind of machines — computers — that would also suggest new mechanisms to govern individuals. This era of the society of control relies on adaptability, modulation, and deformation in order to best match the desired situation. Deleuze writes:
	Such logic updates a disciplinary approach to undetermined behaviour and enters the realm of control. In his 1992 essay, Gilles Deleuze describes a new kind of era, ushered by a new kind of machines — computers — that would also suggest new mechanisms to govern individuals. This era of the society of control relies on adaptability, modulation, and deformation in order to best match the desired situation. Deleuze writes:

	<blockquote>[...] the different control mechanisms are inseparable variations, forming a system of variable geometry the language of which is ''numerical'' (which doesn’t necessarily mean binary). Enclosures are ''molds'', distinct castings, but controls are a ''modulation'', like a self-deforming cast that will continuously change from one moment to the other, or like a sieve whose mesh will transmute from point to point. (Deleuze 3)</blockquote>		<blockquote>[...] the different control mechanisms are inseparable variations, forming a system of variable geometry the language of which is ''numerical'' (which doesn’t necessarily mean binary). Enclosures are ''molds'', distinct castings, but controls are a ''modulation'', like a self-deforming cast that will continuously change from one moment to the other, or like a sieve whose mesh will transmute from point to point. (Deleuze 3)</blockquote>
Line 94:		Line 93:
	Critically inspecting the two educational logics at play in the shaping of vectors — benchmarking as discipline, reinforcement learning as control — highlights two concerns. First, the harmonization of acceptability standards through benchmarks determines the narrow kinds of intelligence which can be expected when interacting with models (i.e. scholarly, academic, bookworm-ish, test-oriented, to the extent that some researchers have started to look into ways to prevent LLMs from cheating on tests (Zhou et al)). For instance, as of 2024, LLMs tend to perform relatively poorly on non-verbal reasoning (Potter). Since the passing of those assessments operate as a sort of test, we can subsequently anticipate ''the kind'' of intelligence that those models display based on their assessment techniques. Second, the fine-tuning of acceptability through reinforcement learning takes a performing academic model resulting from the passing of benchmarks, and presents to the end-user a refined version with particular values embedded in them. Due to the limited amount of companies being able to deploy such reinforcement learning, these values then have similar consideration across the globe (Awad et al). Not only is the factual intelligence standardized, but the values ascribed to those facts is equally controlled.		Critically inspecting the two educational logics at play in the shaping of vectors — benchmarking as discipline, reinforcement learning as control — highlights two concerns. First, the harmonization of acceptability standards through benchmarks determines the narrow kinds of intelligence which can be expected when interacting with models (i.e. scholarly, academic, bookworm-ish, test-oriented, to the extent that some researchers have started to look into ways to prevent LLMs from cheating on tests (Zhou et al)). For instance, as of 2024, LLMs tend to perform relatively poorly on non-verbal reasoning (Potter). Since the passing of those assessments operate as a sort of test, we can subsequently anticipate ''the kind'' of intelligence that those models display based on their assessment techniques. Second, the fine-tuning of acceptability through reinforcement learning takes a performing academic model resulting from the passing of benchmarks, and presents to the end-user a refined version with particular values embedded in them. Due to the limited amount of companies being able to deploy such reinforcement learning, these values then have similar consideration across the globe (Awad et al). Not only is the factual intelligence standardized, but the values ascribed to those facts is equally controlled.

	== Shaping ~~users~~ ==		== Shaping Users ==
	Deleuze’s conception of continuous education as the on-going shaping of intelligences and abilities implies that the structural distinction between what is inside the enclosure and what remains outside is blurry at best. According to the logic of control, the shaping of LLMs does not stop before release to the public. Continuous, user-provided feedback and software updates constantly re-shape their word embeddings (Gao). This last section investigates the potentially shifting positions of tested and tester once LLMs are deployed to — and interacting with — a broader audience.		Deleuze’s conception of continuous education as the on-going shaping of intelligences and abilities implies that the structural distinction between what is inside the enclosure and what remains outside is blurry at best. According to the logic of control, the shaping of LLMs does not stop before release to the public. Continuous, user-provided feedback and software updates constantly re-shape their word embeddings (Gao). This last section investigates the potentially shifting positions of tested and tester once LLMs are deployed to — and interacting with — a broader audience.

Christianvonand at 06:13, 13 October 2024

2024-10-13T06:13:24Z

← Older revision		Revision as of 06:13, 13 October 2024
Line 138:		Line 138:
	<div class="workscited">		<div class="workscited">

	Awad, Edmond, et al. "The Moral Machine Experiment." ''Nature'', vol. 563, no. 7729, Nov. 2018, pp. 59–64. ~~www.nature.com,~~ https://doi.org/10.1038/s41586-018-0637-6.		Awad, Edmond, et al. "The Moral Machine Experiment." ''Nature'', vol. 563, no. 7729, Nov. 2018, pp. 59–64. https://doi.org/10.1038/s41586-018-0637-6.

	Bachimont, Bruno. "Signes Formels et Computation Numérique: Entre Intuition et Formalisme: Critique de La Raison Computationnelle." Instrumente in ''Kunst Und Wissenschaft - Zur Architektonik Kultureller Grenzen Im 17. Jahrhundert'', edited by H Schramm et al., Walter de Gruyter Verlag, 2004.		Bachimont, Bruno. "Signes Formels et Computation Numérique: Entre Intuition et Formalisme: Critique de La Raison Computationnelle." Instrumente in ''Kunst Und Wissenschaft - Zur Architektonik Kultureller Grenzen Im 17. Jahrhundert'', edited by H Schramm et al., Walter de Gruyter Verlag, 2004.

	Bender, Emily M., et al. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜." ''Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency'', ACM, 2021, pp. 610–23. ~~DOI.org (Crossref),~~ https://doi.org/10.1145/3442188.3445922.		Bender, Emily M., et al. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜." ''Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency'', ACM, 2021, pp. 610–23. https://doi.org/10.1145/3442188.3445922.

	Bolukbasi, Tolga, et al. "Man Is to Computer Programmer as Woman Is to Homemaker? Debiasing Word Embeddings." arXiv~~:1607.06520, arXiv~~, 21 July 2016~~. arXiv.org~~, https://doi.org/10.48550/arXiv.1607.06520.		Bolukbasi, Tolga, et al. "Man Is to Computer Programmer as Woman Is to Homemaker? Debiasing Word Embeddings." ''arXiv'', 21 July 2016, https://doi.org/10.48550/arXiv.1607.06520.

	Bowman, Sam. ''Intelligence Testing—Asterisk''. https://asteriskmag.com/issues/04/intelligence-testing. Accessed 2 Mar. 2024.		Bowman, Sam. ''Intelligence Testing—Asterisk''. https://asteriskmag.com/issues/04/intelligence-testing. Accessed 2 Mar. 2024.

	Campbell-Kelly, Martin. From Airline Reservations to Sonic the ~~Hedgehog~~ : A History of the Software Industry. Cambridge, Mass. : MIT Press, 2003. Internet Archive, http://archive.org/details/fromairlinereser00mart_0.		Campbell-Kelly, Martin. ''From Airline Reservations to Sonic the Hedgehog: A History of the Software Industry.'' Cambridge, Mass. : MIT Press, 2003. Internet Archive, http://archive.org/details/fromairlinereser00mart_0.

	Campolo, Alexander, and Katia Schwerzmann. "From Rules to Examples: Machine Learning’s Type of Authority." ''Big Data & Society'', vol. 10, no. 2, July 2023, p. 20539517231188725. SAGE Journals, https://doi.org/10.1177/20539517231188725.		Campolo, Alexander, and Katia Schwerzmann. "From Rules to Examples: Machine Learning’s Type of Authority." ''Big Data & Society'', vol. 10, no. 2, July 2023, p. 20539517231188725. SAGE Journals, https://doi.org/10.1177/20539517231188725.

Simoon at 12:25, 19 September 2024

2024-09-19T12:25:44Z

← Older revision		Revision as of 12:25, 19 September 2024
Line 1:		Line 1:
			'''Pierre Depaz '''
	__NOTOC__		__NOTOC__
	~~'''Pierre Depaz '''~~

	<span class="running-header">Shaping Vectors</span>		<span class="running-header">Shaping Vectors</span>

Geoffcox at 13:42, 18 September 2024

2024-09-18T13:42:50Z

Show changes

Geoffcox at 12:57, 7 September 2024

2024-09-07T12:57:42Z

← Older revision		Revision as of 12:57, 7 September 2024
Line 2:		Line 2:
	'''Pierre Depaz '''		'''Pierre Depaz '''

	<span class="running-header">Shaping Vectors~~: Discipline and Control in Word Embeddings~~</span>		<span class="running-header">Shaping Vectors</span>

	= Shaping Vectors =		= Shaping Vectors =

Simoon at 11:48, 7 September 2024

2024-09-07T11:48:42Z

← Older revision		Revision as of 11:48, 7 September 2024
Line 1:		Line 1:
	__NOTOC__		__NOTOC__
	'''Pierre Depaz '''		'''Pierre Depaz '''

			<span class="running-header">Shaping Vectors: Discipline and Control in Word Embeddings</span>

	= Shaping Vectors =		= Shaping Vectors =

Simoon at 10:16, 7 September 2024

2024-09-07T10:16:07Z

← Older revision		Revision as of 10:16, 7 September 2024
Line 1:		Line 1:
			__NOTOC__
	'''Pierre Depaz '''		'''Pierre Depaz '''
	~~__NOTOC__~~

	= Shaping Vectors =		= Shaping Vectors =
Line 8:		Line 8:

	This article investigates how the word embeddings at the heart of large language models are shaped into acceptable meanings. We show how such shaping follows two educational logics. The use of benchmarks to discover the capabilities of large language models exhibit similar features to Foucault’s disciplining school enclosures, while the process of reinforcement learning is framed as a modulation made explicit in Deleuze’s control societies. The consequences of this shaping into acceptable meaning is argued to result in semantic subspaces. These semantic subspaces are presented as the restricted lexical possibilities of human-machine dialogic interaction, and their consequences are discussed.		This article investigates how the word embeddings at the heart of large language models are shaped into acceptable meanings. We show how such shaping follows two educational logics. The use of benchmarks to discover the capabilities of large language models exhibit similar features to Foucault’s disciplining school enclosures, while the process of reinforcement learning is framed as a modulation made explicit in Deleuze’s control societies. The consequences of this shaping into acceptable meaning is argued to result in semantic subspaces. These semantic subspaces are presented as the restricted lexical possibilities of human-machine dialogic interaction, and their consequences are discussed.

			<div class="page-break"></div>

	== Introduction ==		== Introduction ==

Simoon at 08:39, 5 September 2024

2024-09-05T08:39:02Z

← Older revision		Revision as of 08:39, 5 September 2024
Line 1:		Line 1:
			'''Pierre Depaz '''
	__NOTOC__		__NOTOC__
	~~'''Pierre Depaz '''~~

	= Shaping Vectors =		= Shaping Vectors =