This week’s news of Dr. Michael Burry’s billion-dollar bet against AI-linked equities and the winding down of his hedge fund (https://on.ft.com/43v33a7) arrived at the same time AI pioneer Yann LeCun announced his departure from Meta (https://on.ft.com/4nTkdW7), raising fresh questions about how technology will continue to reshape markets. With credit default swaps on the biggest AI spenders now spiking and trading at stress levels that echo the early rumblings before the 2008 collapse, Burry’s warning that his valuation models are no longer “in sync” with market pricing reflects a growing disconnect between fundamentals and market behaviour. Yet LeCun argues that today’s AI investment frenzy is still chasing a dead end, insisting that real progress will come only from building world-model systems capable of perception and causal reasoning – rather than simply scaling LLMs (https://on.ft.com/43zSWRo).

In markets, LLMs are still mostly used for sentiment extraction, news analysis, research summarisation, and decision support. The engines of signal generation, execution, and risk remain rooted in statistical models, algo trading, and reinforcement learning. Yet trading behaviour is increasingly driven not by fundamentals but by momentum, liquidity, and volatility – echoing Michael Burry’s concerns. Investment decisions today often hinge more on candlestick graphs than on face-to-face CEO meetings. Tools like RavenPack’s Q&A Linguistic Transparency Gap, which uses earnings-call clarity as a signal of future performance (https://www.ravenpack.com/research/qa-language-transparency) and Dave Wangs prompts on social media feeds reflect the speed of shift toward the AI informational edge (https://www.davewang.ai/blog/google-previewed-gemini-3-i-used-ai-to-separate-signal-from-noise).

Both the market’s perceived mispricing of AI and AI’s current technological limitations in trading point to the same conclusion: traders need new frameworks to interpret an environment where information, liquidity, and technology are evolving faster than fundamentals. In a landscape defined by rising information asymmetry, effective trading now demands new skills, new technologies, and new tools suited to a multi-asset, multi-participant, always-on market.

These ideas surfaced directly in conversations this week when I had the privilege of speaking at ‘@Liquidnet’s Gateway, hosted by Jeff Schwartzman, where the next generation of traders met to discuss the future of markets. The themes discussed centred on AI and what it means for their careers:

· How can AI actually be used in trading?

· Is it safe, and what should I be looking for?

· How do I access it with the tools on my desk today?

· How will trading change — will I still have a job?

· Ultimately: how do I make AI work for me?

 Here’s what I learned this week about the developments shaping AI in trading:

1. Building the AI Model: Making LLMs Repeatable

LLMs are inherently probabilistic limiting the ability to implement in trading, but a recent paper from IBM claims that full repeatability is achievable when models are paired with deterministic retrieval and smaller, simpler architectures. The researchers found that smaller models produced identical outputs vs larger models still introduced subtle variation. Most instability came from retrieval rather than generation; once retrieval was made deterministic, lightweight models became fully consistent. This suggests determinism depends not only on settings but on architecture, with smaller models far better suited to stable behaviour. Read more here – https://arxiv.org/pdf/2511.07585

Why this matters for trading: Reliable trading systems require reproducibility, auditability and traceable decision logic – qualities non-deterministic LLMs can easily break. If a model delivers different signals under identical conditions, a desk cannot backtest, justify trades to regulators or maintain stable risk controls. In a domain where decisions must be defensible and consistent, deterministic LLM behaviour is essential for trustworthy, regulation-aligned trading workflows. Thanks to @NickIdelson for sharing.

2. From LLMs to Multi-Agent Systems

While LLMs still remain the core reasoning engines, they need to rely on agent-to-agent coordination through protocols like A2A (https://a2a-protocol.org/latest/)  and on seamless access to tools, data, and APIs via MCP https://www.anthropic.com/news/model-context-protocol). As mentioned above, new alternatives such as Pleias’ compact, high-performance “Baguettotron” show that the next edge may come not from larger models but from fast, specialized ones – potentially also trained more cheaply on synthetic data and orchestrated in dense multi-agent networks (https://huggingface.co/PleIAs/Baguettotron). Add in new efficiency tools like TOON (https://github.com/toon-format/toon) – a token-efficient alternative to JSON – and it’s clear that the challenge ahead is less about LLM or SLM and more about making an entire AI ecosystem operate cleanly, predictably, and at production scale.

Why this matters for trading: Firms that master interoperable layers – MCP for tool access, A2A for agent coordination, TOON for compact data, FIX for execution – are positioned to build true multi-agent trading workflows. In these systems, AI can fetch market data, trigger specialised models, escalate exceptions and supervise risk in real time, at lower latency and cost. As lightweight models become cheaper to train and easier to deploy, competitive advantage shifts to those who integrate them into compliant, end-to-end trading operations.

3. From Terminals to Prompts: Rewriting the Tech Stack

Innovation that once came from sell-side fintech vendors is now emerging from outside the industry. Platforms like https://bigdata.com/, https://www.perplexity.ai/ are part of a fast-growing ecosystem of open-source research agents are reshaping what “institutional research” looks like. Google Finance’s new Deep Search now delivers something close to a Bloomberg-grade research experience for free, enabled by financial rails that can finally support real-time, AI-native workflows (https://blog.google/products/search/new-google-finance-ai-deep-search/) . RavenPack has also open-sourced production-ready research agents (https://www.ravenpack.com/insights/research) while new startup’s like Boltzbit are pushing real-time learning systems that behave less like static models and more like adaptive trading counterparts (https://www.linkedin.com/company/boltzbit/about/). Even day traders now have access to free, high-quality trading prompts online (https://www.davewang.ai/).

Why this matters for trading: The buy-side no longer needs to wait for incumbents to modernise legacy systems. Firms can assemble their own AI-driven stack using open-source technology, compressing the gap between institutional and retail capability, accelerating idea generation and lowering the cost of exploratory research. Incumbent platforms must choose whether to integrate or compete with these tools – or risk being sidelined as traders build flexible, AI-native workflows on their own.

4. What This Means for Traders: The Human-in-the-Loop Decade

The idea that AI is about to erase professions misses the real story. We’re in a multi-year transition where the nature of work is changing faster than the jobs themselves. AI code-generation tools sparked a massive prototyping boom that has now reversed. Usage has dropped 60–80% as teams hit the “complexity wall”: AI can generate code, but it can’t manage integration, security, state, compliance or reliability. Companies are now paying to rebuild and stabilise systems rushed out by AI. As Eduardo Ordax notes, developers aren’t disappearing – they’re cleaning up AI’s output and getting paid well for it (https://www.linkedin.com/feed/update/urn:li:activity:7393503679582023681/).

Why this matters for trading: The question isn’t “Will AI take my job?” but “How will my job change?” AI won’t eliminate trading roles, but it will redefine them. As electronic trading turned voice intermediaries into system-driven strategists, autonomous agents will shift the role again: from executing trades to supervising AI workflows, validating model outputs, managing exceptions and designing guardrails for automated decision-making under uncertainty. Humans won’t be removed from the loop- they will become the supervisors, risk managers and amplifiers of an increasingly automated ecosystem.

5. What This Means for Markets: Resilience and a New Definition of Risk

Anthropic recently reported a cyber-espionage campaign attributed to the state-sponsored Chinese group GTG-1002, notable for its heavy use of AI – specifically Claude Code – to automate the majority of its intrusion workflow (assets.anthropic.com). AI was used for reconnaissance, vulnerability scanning, exploitation, lateral movement, credential harvesting, data analysis and exfiltration, demonstrating a new level of operational scale. Anthropic banned the responsible accounts, notified affected organisations and strengthened detection systems, while calling for greater investment in AI-enabled defence. In parallel, the UK government introduced the Cyber Security and Resilience Bill, expanding regulatory oversight to data centres, MSPs and major digital suppliers (https://www.gov.uk/government/publications/deleted-cyber-security-and-resilience-network-and-information-systems-bill-factsheets/summary-of-the-bill)

Why this matters for trading: As markets become more automated and interconnected, this incident illustrates how autonomous agents can amplify cyber risk if not rigorously secured. Powerful models introduce adaptive, hard-to-predict failure modes, meaning security can no longer be a final-stage control. For trading firms reliant on dense vendor ecosystems, operational, market, systemic and third-party risks converge into a single adaptive risk surface. Managing this requires moving beyond frameworks designed for static threats toward approaches capable of handling dynamic, self-modifying systems. AI can improve prediction of extreme events – but only if risk management evolves.

AI is no longer a niche topic in markets. Its influence is growing, structural and global, introducing new forms of risk, concentration and uncertainty. For trading, this means fresher thinking: different tools, tighter risk controls, shorter time-horizons and far more scenario planning.

Thank you for reading. As always, I’d love to hear your thoughts – what you agree with, where you disagree and what you’d like to see more of next time.

Best wishes,

Rebecca