From Newsgroup: comp.lang.prolog

Hi,

Now there is the neuro symbolic hybrid of tripple store and
artificial neural networks. Already in 2019 proposed an
embedded attention mechanism by Deepak Nathani et al.:

GraphMERT: Efficient and Scalable Distillation
of Reliable Knowledge Graphs from Unstructured Data https://www.researchgate.net/publication/396457862

Neurosymbolic 80M AI from Princeton beats GPT,
SuperIntelligence without OpenAI:
https://www.youtube.com/watch?v=xh6R2WR49yM

Have Fun!

Bye
--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Why do we not see recent Gemini at top here:

"Aider excels with LLMs skilled at writing
and editing code, and uses benchmarks to
evaluate an LLM’s ability to follow instructions
and edit code successfully without human intervention" https://aider.chat/docs/leaderboards/

Google has not yet released a publicly available,
similarly specialized "coding-optimized" version
of Gemini in the same way. The base Gemini models
are generalists, which puts them at a disadvantage
against a fine-tuned specialist in its domain.

There's a concept in the AI community that Google
might be holding back its most capable models
for internal use or specific, controlled

releases (a "moated" approach). A public GraphBERT
by princeton, when it would excell in the coding
domain, might only exercise limited pressure:

- The Trend: The research community has firmly
established that incorporating structural
information (like graphs) improves a model's
understanding of code semantics.

- The Gap: There is currently a performance-efficiency
trade-off. Adding graph processing increases computational
cost and complexity. The largest, most powerful models
(like GPT-4) are so vast that they can implicitly learn
a lot of this structure from a colossal amount of data,
potentially making the explicit graph component less
necessary for them.

Bye

P.S.: Concerning more structured approaches,
somebody shared this paper, quite enjoyable:

*A similar idea*
Getting from Generative AI to Trustworthy AI:
What LLMs might learn from Cyc

Doug Lenat Gary Marcus July 31, 2023
https://arxiv.org/pdf/2308.04445

Mild Shock schrieb:

Hi,

Now there is the neuro symbolic hybrid of tripple store and
artificial neural networks. Already in 2019 proposed an
embedded attention mechanism by Deepak Nathani et al.:

GraphMERT: Efficient and Scalable Distillation
of Reliable Knowledge Graphs from Unstructured Data https://www.researchgate.net/publication/396457862

Neurosymbolic 80M AI from Princeton beats GPT,
SuperIntelligence without OpenAI:
https://www.youtube.com/watch?v=xh6R2WR49yM

Have Fun!

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

Farwell to Martin Kays Translator's Amanuensis.
The old stance for end-user was:

"The goal of MT should not be the fully automatic
high quality translation (FAHQT) that can replace
human translators. Instead, MT should adopt less
ambitious goals, e.g. more cost-effective human-
machine interaction and aim at enhancement of
human translation productivity."

So we now have these spell checkers and grammer
checkes all over the place. But notwithstanding
the glimps of full MT translation in little

browser buttons that could change the language
a full website by automatically translating it.
Something more fundamental happens right now.

Through Generative AI, the copilot not only
is a critic or full translator, he becomes
your pair programmer.

Bye

Mild Shock schrieb:

Hi,

Why do we not see recent Gemini at top here:

"Aider excels with LLMs skilled at writing
and editing code, and uses benchmarks to
evaluate an LLM’s ability to follow instructions
and edit code successfully without human intervention" https://aider.chat/docs/leaderboards/

Google has not yet released a publicly available,
similarly specialized "coding-optimized" version
of Gemini in the same way. The base Gemini models
are generalists, which puts them at a disadvantage
against a fine-tuned specialist in its domain.

There's a concept in the AI community that Google
might be holding back its most capable models
for internal use or specific, controlled

releases (a "moated" approach). A public GraphBERT
by princeton, when it would excell in the coding
domain, might only exercise limited pressure:

 - The Trend: The research community has firmly
   established that incorporating structural
   information (like graphs) improves a model's
   understanding of code semantics.

 - The Gap: There is currently a performance-efficiency
   trade-off. Adding graph processing increases computational
   cost and complexity. The largest, most powerful models
   (like GPT-4) are so vast that they can implicitly learn
   a lot of this structure from a colossal amount of data,
   potentially making the explicit graph component less
   necessary for them.

Bye

P.S.: Concerning more structured approaches,
somebody shared this paper, quite enjoyable:

*A similar idea*
Getting from Generative AI to Trustworthy AI:
What LLMs might learn from Cyc

Doug Lenat  Gary Marcus  July 31, 2023
https://arxiv.org/pdf/2308.04445

Mild Shock schrieb:

Hi,

Now there is the neuro symbolic hybrid of tripple store and
artificial neural networks. Already in 2019 proposed an
embedded attention mechanism by Deepak Nathani et al.:

GraphMERT: Efficient and Scalable Distillation
of Reliable Knowledge Graphs from Unstructured Data
https://www.researchgate.net/publication/396457862

Neurosymbolic 80M AI from Princeton beats GPT,
SuperIntelligence without OpenAI:
https://www.youtube.com/watch?v=xh6R2WR49yM

Have Fun!

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2

From Newsgroup: comp.lang.prolog

Hi,

While ChatGpt is quite forgiving, when we
make spelling mistakes or vague queries, will the
pair programming approach require forgiveness

on our side. It might be the shocking truth, that
this interaction is necessarely a meandering full
of inprecission and errors, but in the same time

converging, performing some mutual goal and knowledge
aquisition. Boris the Loris and Julio the Nazi
Retard were deeply schocked by such an interaction

on SWI-Prolog discourse that even utilized Fuzzy Testing.
Could it be that "communication" is not the core
discipline of Paid Putin Trolls? But "communication"

is the new result of coding:

The New Code — Sean Grove, OpenAI
https://www.youtube.com/watch?v=8rABwKRsec4

and Puting Trolls might eat dust:

Soviet Star Trek - The Final Purge
https://www.youtube.com/watch?v=OtuI2TxFfBA

Bye

Mild Shock schrieb:

Hi,

Farwell to Martin Kays Translator's Amanuensis.
The old stance for end-user was:

"The goal of MT should not be the fully automatic
high quality translation (FAHQT) that can replace
human translators. Instead, MT should adopt less
ambitious goals, e.g. more cost-effective human-
machine interaction and aim at enhancement of
human translation productivity."

So we now have these spell checkers and grammer
checkes all over the place. But notwithstanding
the glimps of full MT translation in little

browser buttons that could change the language
a full website by automatically translating it.
Something more fundamental happens right now.

Through Generative AI, the copilot not only
is a critic or full translator, he becomes
your pair programmer.

Bye

Mild Shock schrieb:

Hi,

Why do we not see recent Gemini at top here:

"Aider excels with LLMs skilled at writing
and editing code, and uses benchmarks to
evaluate an LLM’s ability to follow instructions
and edit code successfully without human intervention"
https://aider.chat/docs/leaderboards/

Google has not yet released a publicly available,
similarly specialized "coding-optimized" version
of Gemini in the same way. The base Gemini models
are generalists, which puts them at a disadvantage
against a fine-tuned specialist in its domain.

There's a concept in the AI community that Google
might be holding back its most capable models
for internal use or specific, controlled

releases (a "moated" approach). A public GraphBERT
by princeton, when it would excell in the coding
domain, might only exercise limited pressure:

  - The Trend: The research community has firmly
    established that incorporating structural
    information (like graphs) improves a model's
    understanding of code semantics.

  - The Gap: There is currently a performance-efficiency
    trade-off. Adding graph processing increases computational
    cost and complexity. The largest, most powerful models
    (like GPT-4) are so vast that they can implicitly learn
    a lot of this structure from a colossal amount of data,
    potentially making the explicit graph component less
    necessary for them.

Bye

P.S.: Concerning more structured approaches,
somebody shared this paper, quite enjoyable:

*A similar idea*
Getting from Generative AI to Trustworthy AI:
What LLMs might learn from Cyc

Doug Lenat  Gary Marcus  July 31, 2023
https://arxiv.org/pdf/2308.04445

Mild Shock schrieb:

Hi,

Now there is the neuro symbolic hybrid of tripple store and
artificial neural networks. Already in 2019 proposed an
embedded attention mechanism by Deepak Nathani et al.:

GraphMERT: Efficient and Scalable Distillation
of Reliable Knowledge Graphs from Unstructured Data
https://www.researchgate.net/publication/396457862

Neurosymbolic 80M AI from Princeton beats GPT,
SuperIntelligence without OpenAI:
https://www.youtube.com/watch?v=xh6R2WR49yM

Have Fun!

Bye

--- Synchronet 3.21a-Linux NewsLink 1.2