Ciaran O'Reilly

[Ciaran O'Reilly @ccoreilly]

@parloa_ai has closed its Series D funding round and raised $350M at a $3 billion valuation 🚀

What an incredible achievement for our entire team!

Learn more about our Series D here: https://t.co/8NEOncXJX9

[Projecte 'Ce Trencada' @PrCeTrencada]

🟤 Recordeu que avui, 31/07/25, es llança arreu el #videojoc anomenat #NinjaGaidenRagebound, disponible oficialment en #català gràcies al @DavidJaumandreu! Moltes gràcies per pensar en la nostra cultura en un llançament tan important. Quin autèntic goig!👉 https://t.co/j8nlrrIYrO

[Pep Toni @PepToni25]

🧵Demà sortirà el videojoc Ninja Gaiden Ragebound i avui s'han publicat les notes i ressenyes de la premsa.

Però el que té d'especial aquest videojoc és que té traducció en català. I aquest fil precisament és per parlar sobre el ressò que s'ha fet d'aquesta traducció 👇

[Softcatalà @softcatala]

Presentem un servei de doblatge automàtic de vídeos en català en proves

Vos animem a provar-lo i a ajudar-nos a millorar-lo 🫶🏻

https://t.co/kbDTjNtd4K

[Parloa @parloa_ai]

The shift from conversational AI to agentic AI in customer service also introduces new risk and security challenges. As autonomous AI agents evolve, businesses must develop new strategies to win customer trust.

To help businesses in this new era, @opusresearch just released an extensive report on genAI trust and safety. It includes some innovative approaches that Parloa has built to mitigate risks with genAI.

For example, our AI Agent Management Platform provides quality assurance (QA) through large-scale simulated conversations across multiple test scenarios, to ensure AI agents are ready for real-world conversations — before ever talking to customers.

Simulation-based testing is a critical component in any enterprise-grade AI security strategy.

To learn more how enterprises should mitigate risks when adopting agentic AI, download your copy of Opus Research’s latest report 👇https://t.co/1ZjDnWY6Ma

#Parloa #CustomerService #AgenticAI #AIAgents

[Andrej Karpathy @karpathy]

# RLHF is just barely RL

Reinforcement Learning from Human Feedback (RLHF) is the third (and last) major stage of training an LLM, after pretraining and supervised finetuning (SFT). My rant on RLHF is that it is just barely RL, in a way that I think is not too widely appreciated. RL is powerful. RLHF is not. Let's take a look at the example of AlphaGo. AlphaGo was trained with actual RL. The computer played games of Go and trained on rollouts that maximized the reward function (winning the game), eventually surpassing the best human players at Go. AlphaGo was not trained with RLHF. If it were, it would not have worked nearly as well.

What would it look like to train AlphaGo with RLHF? Well first, you'd give human labelers two board states from Go, and ask them which one they like better:

Then you'd collect say 100,000 comparisons like this, and you'd train a "Reward Model" (RM) neural network to imitate this human "vibe check" of the board state. You'd train it to agree with the human judgement on average. Once we have a Reward Model vibe check, you run RL with respect to it, learning to play the moves that lead to good vibes. Clearly, this would not have led anywhere too interesting in Go. There are two fundamental, separate reasons for this:

1. The vibes could be misleading - this is not the actual reward (winning the game). This is a crappy proxy objective. But much worse,
2. You'd find that your RL optimization goes off rails as it quickly discovers board states that are adversarial examples to the Reward Model. Remember the RM is a massive neural net with billions of parameters imitating the vibe. There are board states are "out of distribution" to its training data, which are not actually good states, yet by chance they get a very high reward from the RM.

For the exact same reasons, sometimes I'm a bit surprised RLHF works for LLMs at all. The RM we train for LLMs is just a vibe check in the exact same way. It gives high scores to the kinds of assistant responses that human raters statistically seem to like. It's not the "actual" objective of correctly solving problems, it's a proxy objective of what looks good to humans. Second, you can't even run RLHF for too long because your model quickly learns to respond in ways that game the reward model. These predictions can look really weird, e.g. you'll see that your LLM Assistant starts to respond with something non-sensical like "The the the the the the" to many prompts. Which looks ridiculous to you but then you look at the RM vibe check and see that for some reason the RM thinks these look excellent. Your LLM found an adversarial example. It's out of domain w.r.t. the RM's training data, in an undefined territory. Yes you can mitigate this by repeatedly adding these specific examples into the training set, but you'll find other adversarial examples next time around. For this reason, you can't even run RLHF for too many steps of optimization. You do a few hundred/thousand steps and then you have to call it because your optimization will start to game the RM. This is not RL like AlphaGo was.

And yet, RLHF is a net helpful step of building an LLM Assistant. I think there's a few subtle reasons but my favorite one to point to is that through it, the LLM Assistant benefits from the generator-discriminator gap. That is, for many problem types, it is a significantly easier task for a human labeler to select the best of few candidate answers, instead of writing the ideal answer from scratch. A good example is a prompt like "Generate a poem about paperclips" or something like that. An average human labeler will struggle to write a good poem from scratch as an SFT example, but they could select a good looking poem given a few candidates. So RLHF is a kind of way to benefit from this gap of "easiness" of human supervision. There's a few other reasons, e.g. RLHF is also helpful in mitigating hallucinations because if the RM is a strong enough model to catch the LLM making stuff up during training, it can learn to penalize this with a low reward, teaching the model an aversion to risking factual knowledge when it's not sure. But a satisfying treatment of hallucinations and their mitigations is a whole different post so I digress. All to say that RLHF *is* net useful, but it's not RL.

No production-grade *actual* RL on an LLM has so far been convincingly achieved and demonstrated in an open domain, at scale. And intuitively, this is because getting actual rewards (i.e. the equivalent of win the game) is really difficult in the open-ended problem solving tasks. It's all fun and games in a closed, game-like environment like Go where the dynamics are constrained and the reward function is cheap to evaluate and impossible to game. But how do you give an objective reward for summarizing an article? Or answering a slightly ambiguous question about some pip install issue? Or telling a joke? Or re-writing some Java code to Python? Going towards this is not in principle impossible but it's also not trivial and it requires some creative thinking. But whoever convincingly cracks this problem will be able to run actual RL. The kind of RL that led to AlphaGo beating humans in Go. Except this LLM would have a real shot of beating humans in open-domain problem solving.

[Apoorv Agrawal @apoorv03]

What would AI x customer service look like if we built it from scratch today?

If you are a large enterprise who is answering this question, reach out to @parloa_ai to learn about their AI Agent Platform

@maltekosub

[Ciaran O'Reilly @ccoreilly]

Really proud of what we have been building in the last months at @parloa_ai

[Parloa @parloa_ai]

Today, we’re announcing a new chapter for #AI in customer service. 🚀

Imagine a personal AI agent for every customer.

In the #genAI era, if you have millions of customers, you should be having millions of unique conversations.

This is why we’re introducing a new category of software, to empower enterprises to connect with every single one of their customers, safely and at scale:

✨ The Parloa AI Agent Management Platform ✨

Or simply, Parloa AMP.

▫️ Parloa AMP allows companies to provide personal AI agents for every customer. They are designed with natural language briefings, not scripted flows, so they can have dynamic conversations across a complex set of use cases — and faster time to value.

▫️ Parloa AMP provides a portfolio of agent lifecycle management tools that allows companies to design, QA, deploy, and scale a team of personal AI agents. It can simulate and evaluate thousands of synthetic conversations before deploying to customers so you can fine-tune each agent.

Our co-founders @maltekosub and @o_stefan will unveil AMP live on stage on September 12th at our WAVE conference in #Berlin, join us!

In the meantime, here's more on today’s sneak peek: https://t.co/IdLohve8NU

[Parloa @parloa_ai]

Today, we’re announcing a new chapter for #AI in customer service. 🚀

Imagine a personal AI agent for every customer.

In the #genAI era, if you have millions of customers, you should be having millions of unique conversations.

This is why we’re introducing a new category of software, to empower enterprises to connect with every single one of their customers, safely and at scale:

✨ The Parloa AI Agent Management Platform ✨

Or simply, Parloa AMP.

▫️ Parloa AMP allows companies to provide personal AI agents for every customer. They are designed with natural language briefings, not scripted flows, so they can have dynamic conversations across a complex set of use cases — and faster time to value.

▫️ Parloa AMP provides a portfolio of agent lifecycle management tools that allows companies to design, QA, deploy, and scale a team of personal AI agents. It can simulate and evaluate thousands of synthetic conversations before deploying to customers so you can fine-tune each agent.

Our co-founders @maltekosub and @o_stefan will unveil AMP live on stage on September 12th at our WAVE conference in #Berlin, join us!

In the meantime, here's more on today’s sneak peek: https://t.co/IdLohve8NU

[Apoorv Agrawal @apoorv03]

Customer service will be redefined with AI.

Fierce battle brewing in the space, but Parloa stands out as one focused on the largest enterprises, with the highest quality product across voice/text etc.

@maltekosub and I dig into our partnership below CC @altcap

[Mirco Ravanelli @mirco_ravanelli]

Exciting news! 🎉 #SpeechBrain 1.0 is out with tons of thrilling advancements.

Our #OpenSource toolkit now features 200+ recipes and 100+ pretrained models on #HuggingFace for diverse #ConversationalAI tasks.

🌐 Website: https://t.co/a1wqxLucgw

💻 Repo: https://t.co/MsCZbSbSOf

[Ciaran O'Reilly @ccoreilly]

@_josh_meyer_ Really sad to read this. The great community around DeepSpeech helped me get into STT. It further strived in the matrix servers when you founded Coqui and managed to build great stuff for TTS. I wish you, @erogol , @ReubenMorais and the rest of the team all the best. 💙

[Ciaran O'Reilly @ccoreilly]

@xaviviro @projecte_aina @teknium @cibernicola_es Gran feina Xavi! Els conjunts de dades i scripts per fer el fine-tuning els estàs publicant a algun lloc? Seria molt útil per qui tingui interès en aprendre de tu!

[Xavi Vinaixa @xaviviro]

Acabo de publicar FLAMA el primer model petit (3B) #LLM en català i preparat pel format ChatML. També en format GGUF i quantitzat. #ia #iaencatalà #llm #llmencatalà #català #encatalà

👉🏻 https://t.co/k2rq6gR5B4

[Jordi Mas @jordimash]

Avui fem els #25anysSoftcatalà. Com passa el temps.

Ens vam presentar un 2 d’octubre de 1998, però feia mesos que fèiem feina.

Aquí hi ha la història d'aquells primers temps:
https://t.co/6vMAcb6cOl

Aquests dies parlaré de passat, present futur. Comencem amb el passat 🧵👇

[Atila @atiorh]

Stable Diffusion XL with Core ML on Apple Silicon! #SDXL The model grew 3x in size to ~2.6 billion parameters so we are releasing a new model compression technique that yields variants quantized to as little as 3 bits with minimal output difference 🧵

[Marta Villegas @MartaVillegasM]

📢Aguila-7B, a new open source transformer language model for Spanish, Catalan & English. Based on the Falcon-7B model, it has been trained on a 26B token trilingual corpus collected from available corpora.
🚀https://t.co/mkpfaye2Tj
@projecte_aina @SEDIAgob @tic @BSC_CNS @PLNnet

[Jordi Mas @jordimash]

Google ha presentat AudioPaLM. Combina models de llengua grans amb models d'àudio. Permet fer traducció veu a veu (conversa) i reconeixement de veu. El català té suport de primer nivell (ho consideren llengua «high resource»), i de fet surt als exemples.

https://t.co/44TP8575z3

[Joan Montané 🏴󠁧󠁢󠁥󠁮󠁧󠁿🌹📖🐉 @unjoanqualsevol]

Avui el #català és la llengua amb més hores enregistrades al @mozilla #CommonVoice. Superant l'anglès. 3248 hores! 🤯

Sense l'impuls de moltes entitats i de les més de 34.000 persones que han donat la seva veu durant gairebé 5 anys, no hauria estat possible. 😘

#CommonVoiceCAT