Connect with us
DAPA Banner

Crypto World

Lessons Learned After a Year of Building with Large Language Models (LLMs)

Published

on

Lessons Learned After a Year of Building with Large Language Models (LLMs)

Over the past year, Large Language Models (LLMs) have reached impressive competence for real-world applications. Their performance continues to improve, and costs are decreasing, with a projected $200 billion investment in artificial intelligence by 2025. Accessibility through provider APIs has democratised access to these technologies, enabling ML engineers, scientists, and anyone to integrate intelligence into their products. However, despite the lowered entry barriers, creating effective products with LLMs remains a significant challenge. This is summary of the original paper of the same name by https://applied-llms.org/. Please refer to that documento for detailed information.

Fundamental Aspects of Working with LLMs


· Prompting Techniques


Prompting is one of the most critical techniques when working with LLMs, and it is essential for prototyping new applications. Although often underestimated, correct prompt engineering can be highly effective.

– Fundamental Techniques: Use methods like n-shot prompts, in-context learning, and chain-of-thought to enhance response quality. N-shot prompts should be representative and varied, and chain-of-thought should be clear to reduce hallucinations and improve user confidence.

Structuring Inputs and Outputs: Structured inputs and outputs facilitate integration with subsequent systems and enhance clarity. Serialisation formats and structured schemas help the model better understand the information.

Advertisement

– Simplicity in Prompts: Prompts should be clear and concise. Breaking down complex prompts into more straightforward steps can aid in iteration and evaluation.

– Token Context: It’s crucial to optimise the amount of context sent to the model, removing redundant information and improving structure for clearer understanding.


· Retrieval-Augmented Generation (RAG)


RAG is a technique that enhances LLM performance by providing additional context by retrieving relevant documents.


– Quality of Retrieved Documents: The relevance and detail of the retrieved documents impact output quality. Use metrics such as Mean Reciprocal Rank (MRR) and Normalised Discounted Cumulative Gain (NDCG) to assess quality.

Advertisement

– Use of Keyword Search: Although vector embeddings are useful, keyword search remains relevant for specific queries and is more interpretable.

– Advantages of RAG over Fine-Tuning: RAG is more cost-effective and easier to maintain than fine-tuning, offering more precise control over retrieved documents and avoiding information overload.


Optimising and Tuning Workflows


Optimising workflows with LLMs involves refining and adapting strategies to ensure efficiency and effectiveness. Here are some key strategies:


· Step-by-Step, Multi-Turn Flows


Decomposing complex tasks into manageable steps often yields better results, allowing for more controlled and iterative refinement.

Advertisement

– Best Practices: Ensure each step has a defined goal, use structured outputs to facilitate integration, incorporate a planning phase with predefined options, and validate plans. Experimenting with task architectures, such as linear chains or Directed Acyclic Graphs (DAGs), can optimise performance.


· Prioritising Deterministic Workflows


Ensuring predictable outcomes is crucial for reliability. Use deterministic plans to achieve more consistent results.

Benefits: It facilitates controlled and reproducible results, makes tracing and fixing specific failures easier, and DAGs adapt better to new situations than static prompts.

– Approach: Start with general objectives and develop a plan. Execute the plan in a structured manner and use the generated plans for few-shot learning or fine-tuning.

Advertisement


· Enhancing Output Diversity Beyond Temperature


Increasing temperature can introduce diversity but only sometimes guarantees a good distribution of outputs. Use additional strategies to improve variety.

– Strategies: Modify prompt elements such as item order, maintain a list of recent outputs to avoid repetitions, and use different phrasings to influence output diversity.


· The Underappreciated Value of Caching


Caching is a powerful technique for reducing costs and latency by storing and reusing responses.

– Approach: Use unique identifiers for cacheable items and employ caching techniques similar to search engines.

Advertisement

– Benefits: Reduces costs by avoiding recalculation of responses and serves vetted responses to reduce risks.


· When to Fine-Tune


Fine-tuning may be necessary when prompts alone do not achieve the desired performance. Evaluate the costs and benefits of this technique.

– Examples: Honeycomb improved performance in specific language queries through fine-tuning. Rechat achieved consistent formatting by fine-tuning the model for structured data.

– Considerations: Assess if the cost of fine-tuning justifies the improvement and use synthetic or open-source data to reduce annotation costs.

Advertisement


Evaluation and Monitoring


Effective evaluation and monitoring are crucial to ensuring LLM performance and reliability.

· Assertion-Based Unit Tests


Create unit tests with real input/output examples to verify the model’s accuracy according to specific criteria.

– Approach: Define assertions to validate outputs and verify that the generated code performs as expected.


· LLM-as-Judge

Use an LLM to evaluate the outputs of another LLM. Although imperfect, it can provide valuable insights, especially in pairwise comparisons.

Advertisement

– Best Practices: Compare two outputs to determine which is better, mitigate biases by alternating the order of options and allowing ties, and have the LLM explain its decision to improve evaluation reliability.


· The “Intern Test”

Evaluate whether an average university student could complete the task given the input and context provided to the LLM.

– Approach: If the LLM lacks the necessary knowledge, enrich the context or simplify the task. Decompose complex tasks into simpler components and investigate failure patterns to understand model shortcomings.


· Avoiding Overemphasis on Certain Evaluations

Do not focus excessively on specific evaluations that might distort overall performance metrics.

Advertisement

Example: A needle-in-a-haystack evaluation can help measure recall but does not fully capture real-world performance. Consider practical assessments that reflect real use cases.


Key Takeaways


The lessons learned from building with LLMs underscore the importance of proper prompting techniques, information retrieval strategies, workflow optimisation, and practical evaluation and monitoring methodologies. Applying these principles can significantly enhance your LLM-based applications’ effectiveness, reliability, and efficiency. Stay updated with advancements in LLM technology, continuously refine your approach, and foster a culture of ongoing learning to ensure successful integration and an optimised user experience.

Source link

Advertisement
Continue Reading
Click to comment

You must be logged in to post a comment Login

Leave a Reply

Crypto World

Bitcoin Dips Below $70,000 as Extreme Fear Index Hits 10: What Traders Are Watching Next

Published

on

Brian Armstrong's Bold Prediction: AI Agents Will Soon Dominate Global Financial

TLDR:

  • Bitcoin fell over 3% in 24 hours, sliding from above $74,000 to around $68,700 on Sunday amid macro fears.
  • The Crypto Fear and Greed Index dropped to an extreme fear reading of 10, reflecting sharp decline in market confidence.
  • Trader Lennaert Snyder targets a Bitcoin drop to $65,580, planning to add shorts after a confirmed bearish structure break.
  • Institutional buyers continue accumulating BTC as exchange supply hits multi-year lows, contrasting with heavy retail panic selling.

Bitcoin fell sharply on Sunday, dropping from above $74,000 to around $68,700 in a matter of hours. The move pushed the Crypto Fear & Greed Index to an extreme fear reading of just 10.

Rising oil prices, a pause in Federal Reserve rate cuts, and ongoing geopolitical tensions drove the sell-off. Bitcoin recorded a 3.11% decline over 24 hours, with trading volume reaching approximately $29.1 billion.

Short Positions Build as Bears Set Their Sights on $65,000

The latest price drop has given bearish traders confidence to hold and grow their short positions. Selling pressure remained active throughout the week, contributing to a total seven-day decline of 4.02%.

This combination of macro pressure and bearish momentum pushed market fear to its most extreme reading in recent weeks.

Crypto trader Lennaert Snyder shared his bearish stance openly on social media during Sunday’s session. “My target is still the ~$65,580 low, and possibly even lower for Bitcoin,” Snyder wrote. He also planned to add margin to his shorts using the upper wick of the next weekly candle.

Advertisement

Snyder noted caution around a key level at $72,700, identifying it as a Fair Value Gap zone. He stated he would only enter a trade after seeing a liquidity push and a bearish market structure break.

His approach pointed to a disciplined strategy, waiting for price confirmation before committing to new short trades.

A notable counterrisk, however, remains for those currently holding short positions. Whale Insider reported that $5 billion in crypto shorts would face forced liquidation if Bitcoin climbs back to $75,000. That level therefore becomes both a target for bulls and a danger zone for active short sellers in the market.

Advertisement

Institutional Buyers Accumulate as Exchange Supply Drops to Multi-Year Lows

Even as retail sentiment fell to extreme fear, institutional buyers continued accumulating Bitcoin through the downturn.

This divergence between retail and large-scale buyers has been a repeated pattern during past crypto market corrections. Institutions appear to view the current dip as an entry point rather than a reason to sell.

Exchange supply has also dropped to multi-year lows, further shaping the current market picture. Lower exchange balances typically point to Bitcoin being moved into cold storage for long-term holding.

This movement often tightens available sell-side supply on exchanges, setting the stage for potential price rebounds.

Advertisement

Market watchers are now turning their attention to Monday’s session, closely eyeing the $72,000 price level. A recovery above that zone could signal a momentum shift and place short positions at increased risk. Bulls will need consistent buying volume to challenge the bearish tone that dominated the weekend.

Bitcoin’s near-term path will largely depend on how macro factors unfold over the coming days. Bears are holding firm to the $65,580 target, while bulls look for a sustained break above $72,000.

The market remains at a crossroads, with either outcome carrying major consequences for active traders on both sides.

Advertisement

Source link

Continue Reading

Crypto World

Resolv Labs confirms no loss of assets after USR exploit shakes market

Published

on

Resolv Labs confirms no loss of assets after USR exploit shakes market

Resolv Labs recently experienced a major exploit in its USR stablecoin system, leading to the minting of 80 million unbacked tokens. 

Summary

  • USR stablecoin crashes to $0.14 after exploit, rebounding to $0.42.
  • DeFi protocols quickly respond to exploit, with some pausing markets to limit risk.
  • Resolv Labs reassures users, stating collateral pool remains intact despite exploit.

Meanwhile, this triggered a sharp drop in the token’s value, causing it to fall as low as $0.14 before rebounding to $0.42. The incident has raised concerns among decentralized finance (DeFi) protocols and users exposed to the exploit, prompting a rapid response to contain the fallout.

As Crypto News reported earlier on Sunday, Resolv Labs confirmed that an attacker had exploited the minting mechanics of its USR stablecoin. The attacker was able to create tens of millions of unbacked USR tokens and sell them through DeFi pools. This led to a dramatic depeg of the token, which dropped as low as $0.14, 86% below its intended $1 value.

Advertisement

The price of USR quickly rebounded to $0.42, but the attack had already caused significant damage. Resolv Labs reassured users by stating that the collateral pool “remains fully intact” and that the issue was isolated to the USR issuance mechanics. The team has paused the protocol to assess the situation and prevent further exploitation.

Following the exploit, DeFi protocols that had exposure to USR moved quickly to contain any potential damage. Lido, Morpho, and Aave all issued statements confirming that their systems were unaffected, although some vaults did have exposure to the exploit.

According to Michael Pearl of Cyvers, the risk from the exploit seemed concentrated in lending and leverage markets, particularly those using USR or RLP as collateral. Some platforms like Euler, Venus, and Fluid paused markets or isolated vaults to prevent further risks. Pearl noted that the impact appeared to be localized, with no signs of a broader contagion affecting the entire DeFi ecosystem.

Advertisement

Moreover, despite Resolv Labs’ smart contracts undergoing multiple audits, the exploit has raised questions about the limitations of these audits. Security firm Pashov, which had audited Resolv’s staking module in July 2025, pointed out that the attack likely stemmed from an operational security flaw rather than a design issue. The firm highlighted the potential compromise of a private key as the root cause of the exploit.

Experts like Pearl argued that real-time monitoring powered by artificial intelligence is essential to detect anomalies in protocol activity. Monitoring mint and burn flows and validating supply against reserves would help detect issues before they escalate.

Containment and recovery efforts

Resolv Labs has reassured its users that it is actively investigating the exploit and working on recovery. While the exploit did not result in any loss of assets from the collateral pool, the attack has emphasized the need for continuous monitoring and stronger operational security. The DeFi community is closely watching how Resolv Labs handles the situation, especially as the price of USR stabilizes and more data on the full impact of the exploit becomes available.

Advertisement

Source link

Continue Reading

Crypto World

TSMC Helium Crisis: How the Persian Gulf War Put the World’s Chip Supply on an 11-Day Clock

Published

on

Brian Armstrong's Bold Prediction: AI Agents Will Soon Dominate Global Financial

TLDR:

  • TMSC holds only 11 days of LNG reserve, the least of any major semiconductor economy on Earth.
  • Helium from Qatar powers EUV machines that print every advanced AI chip at 3-nanometre scale globally.
  • Helium spot prices have surged up to 100% since Iranian strikes shut down Qatar’s Ras Laffan complex.
  • Two US carrier strike groups have shifted to the Gulf, thinning Pacific presence and raising Taiwan risk.

TSMC produces 90 percent of the world’s most advanced logic chips. Taiwan, where TSMC operates, imports 97 percent of its energy and holds only 11 days of gas in reserve.

A war in the Persian Gulf has now disrupted Taiwan’s helium supply. Helium is critical for printing transistors at 3 nanometres, with no substitute available. The crisis has put global semiconductor supply chains under immediate pressure.

Helium Shortage Pushes Advanced Chip Manufacturing Toward a Critical Threshold

Qatar’s Ras Laffan complex once processed roughly one-third of the world’s helium. Iranian strikes shut it down, and repairs will take three to five years.

Taiwan relies on Qatar for the bulk of its helium supply. SK Hynix also sourced 64.7 percent of its helium from Qatar. Helium spot prices have since surged between 40 and 100 percent.

Helium cools the EUV lithography systems that print chips at 3 nanometres. It purges etching chambers of contamination and tests wafer seals.

Advertisement

No substitute for helium exists in these manufacturing processes. Without it, EUV machines stop entirely not slowly, but completely.

Analyst Shanaka Perera wrote on X that helium is “the molecule the market is not pricing.” He added that without it, EUV machines stop “not slow down. Stop.” Bloomberg reported TSMC may prioritise AI chip production over consumer products during shortages.

Fitch Ratings flagged Taiwan and South Korea as the most exposed semiconductor economies. TSMC’s shares have fallen 7 percent since the war began.

Advertisement

Taiwan holds the smallest energy reserve among major semiconductor economies. South Korea holds 52 days of reserve; Japan holds three weeks.

Geopolitical Pressure Compounds Taiwan’s Strategic Energy Exposure

Taiwan’s Ministry of Economic Affairs says helium supplies are secured through mid-May. Negotiations for June are ongoing, and officials called the situation a controllable risk. The government also announced plans to raise the mandatory LNG reserve from 11 to 14 days next year.

The Persian Gulf war has redirected two US carrier strike groups away from the Pacific. This has thinned the naval presence that historically deters pressure on Taiwan. Regional tensions around Taiwan have been building since 2023.

Beijing does not need an invasion to apply pressure on Taiwan. A military exercise near the island during a supply crisis achieves disruption through perception. That signal alone can alter market behaviour and shipping logistics.

Advertisement

Perera noted that seven reinsurance letters closed the Strait of Hormuz commercially in five days. The same mechanism could apply to the Taiwan Strait, which is 110 miles wide at its broadest point. If risk models shift, insurance letters follow, and shipping stops without any military action.

Taiwan imports 97 percent of its energy, with one-third from the Middle East. Qatar remains the dominant LNG supplier.

The chain connecting helium, LNG, and the world’s advanced chips now runs through an active war zone. TSMC remains the most critical manufacturer of advanced semiconductors on Earth.

Advertisement

Source link

Continue Reading

Crypto World

Resolv Says No Assets Lost After USR Stablecoin Exploit

Published

on

Cryptocurrencies, Smart Contracts, Hacks, Stablecoin, DeFi

Resolv Labs moved Sunday to reassure users after an exploit hit the issuance mechanics of its USR stablecoin, knocking the token off its dollar peg and prompting decentralized finance (DeFi) protocols with exposure to move quickly to contain any fallout.

Cointelegraph reported earlier Sunday that an attacker exploited USR’s minting mechanics, creating tens of millions of unbacked tokens and dumping them through DeFi pools, which broke the stablecoin’s peg and prompted Resolv to pause protocol functions as it assessed the damage.

The token dropped as low as $0.14 (86% below its intended $1 price) after the exploit before rebounding to $0.42 at the time of writing, according to data from CoinGecko.

In a recent statement on X, the Resolv team said that the collateral pool “remains fully intact,” and that the problem appears “isolated to USR issuance mechanics.” Containment and impact assessment remain ongoing.

Advertisement

Onchain data from Arkham, corroborated by Web3 security firm Cyvers, showed that the attacker had converted most of the minted USR into Ether (ETH), selling part of the haul for about 11,400 ETH (around $24 million). Independent analysts also noted that the remaining 36.74 million USR was “still being continuously dumped.”

Cryptocurrencies, Smart Contracts, Hacks, Stablecoin, DeFi
USR dropped 86% off its peg. Source. CoinGecko

Michael Pearl, vice president GTM and strategy at Cyvers, told Cointelegraph that since the supply had inflated faster than the market could absorb and the token had immediately depegged, the value of the remaining tokens was significantly impaired.

Related: Google Threat Intel flags ‘Ghostblade’ crypto-stealing malware

DeFi protocols move to contain fallout

Decentralized finance (DeFi) protocols with exposure to Resolv raced to clarify their positions. Liquid staking provider Lido said that Lido Earn user funds were safe. Morpho cofounder Merlin Egalite emphasized that the lending protocol’s own contracts were unaffected and that only certain vaults had exposure, and Aave’s founder, Stani Kulechov, said that the platform had no direct USR exposure and that Resolv was repaying its outstanding debt.

The X account “yieldsandmore” pointed to potential losses in Resolv’s junior RLP tranche, highlighting possible knock-on effects for yield platforms such as Stream and yoUSD that used RLP as collateral. 

Advertisement

Pearl told Cointelegraph that, based on available data, the exposure appeared to be “relatively concentrated” in lending markets and leverage loops “rather than system-wide,” and primarily in protocols that integrated USR, wstUSR, or RLP into lending, leverage or yield strategies.

Related: Hacked crypto tokens drop 61% on average and rarely recover, Immunefi report says

He said that several protocols, such as Euler, Venus, Lista and Fluid, had taken precautionary actions such as pausing markets or isolating vaults, while others had declared no exposure at all. “It is more accurate to describe the risk as concentrated with localized spillover, rather than widespread contagion,” he said.

Ledger chief technical officer Charles Guillemet also assessed the fallout on X, stating that, due to the relatively small size of USR, “this is not a Terra Luna-type event.”

Advertisement

Questions around limitations of security audits

Resolv’s smart contracts have undergone multiple audits since 2024, but Pearl said that, while audits were “necessary,” they were also “inherently static and scoped.” Real-time, artificial intelligence-powered monitoring to “continuously analyze protocol activity” was needed, he argued, to detect anomalies as they emerge.

For stablecoin systems specifically, he said that meant monitoring mint and burn flows against expected behavior in real time, continuously validating supply against reserves and backing assets, and detecting anomalies in oracle inputs, pricing and liquidity conditions. 

Security firm Pashov, which audited Resolv’s staking module in July 2025, told Cointelegraph that Resolv’s design was “good,” and that the root cause was “not the design so much as the private key compromise,” which was likely an operational security flaw. “We have to understand how that happens,” he said.

Cointelegraph reached out to Resolv Labs for comment but had not received a response by publication.

Advertisement

AI Eye: IronClaw rivals OpenClaw, Olas launches bots for Polymarket